JP3653804B2 - Particle image region segmentation method and apparatus - Google Patents

Description

但し、ｉ＝０，１，２，…，５１０，５１１、ｊ＝０，１，２，…，５１０，５１１である。
【００５４】
また近似的に、黒画像Ｂ（ｉ，ｊ）の分散が、Ａ／Ｄ変換された値で１レベル以下であれば、黒画像Ｂ（ｉ，ｊ）をその平均値Ｂ_avで置き換え、白画像Ｗ（ｉ，ｊ）も各画素毎を参照することなく、小領域の白画像Ｗ’（ｉ’，ｊ’）での画素を参照するものとすれば、次の（数３）により、上記と同様の補正ができる。
【００５５】

ここで、参照する白画像の小領域の大きさをＭ×Ｎとすると、（ｉ’，ｊ’）は参照する座標を示し、
ｉ’＝［ｉ／Ｍ］
ｊ’＝［ｊ／Ｎ］ …（数４）
であり、［ ］はガウス記号である。例えば、Ｍ＝Ｎ＝８とすれば、白画像Ｗ’の大きさは、６４×６４画素分の大きさでよいことになる。
【００５６】
黒レベルの基準値β₀、白レベルの基準値α₀は、各々Ｒ画像、Ｇ画像、Ｂ画像ともに、β₀＝２３０、α＝４０程度の値に設定する。
【００５７】
また、白画像は、その画像内で平滑化処理を行なうか、あるいは複数の白画像を計測して、画像の平均処理、又は平滑化処理を行ない、ノイズの影響が低減できる。この補正により、撮影された色画像は、画像全面に渡って濃度が、ほぼ均一になる。以上説明した、画像の補正は、光学系の歪みが、実質的に無視できるときには、省略できることは言うまでもない。
【００５８】
以上の説明は、以下の各実施例において、領域分割を行なうに先立って実行される共通する処理である。
【００５９】
（実施例１）
実施例１は、本願発明の構成の一部を示す実施例であり、濃度値に基づいて、背景領域と対象領域とを識別する構成について、具体的に説明する。本実施例では、フロー方式で撮像した画像中に存在する、大きさの異なる（数μｍ〜数百μｍ）複数個の粒子を効率よく分割するために、ＴＶカメラで波長別に入力した画像から、光学系の濃度むらを補正した後に、赤画像、緑画像の色空間で、背景領域と、対象粒子が存在する対象領域を分割し、背景領域と、対象粒子が存在する対象領域を表わす２値画像を得る。
【００６０】
本実施例では、検査対象粒子の自動識別に好適な、青と赤のコントラストが高いＳ染色で染色した試料の画像から、画像中の沈渣成分の領域を分割する。領域分割は、画像の背景領域を基準として行なう。予め上記で説明した補正処理を行ない、光学系の濃度むらを補正し、濃度変動の偏差を小さくして、背景領域の安定な抽出を行なう。
【００６１】
以下、領域分割方法の詳細について説明する。
【００６２】
以下に説明する領域分割方法では、Ｓ染色で染色された試料の画像から、目的とする対象粒子の対象領域の抽出に際し、濃度差のある粒子像や色調差のある粒子像でも、安定に分割できる特徴がある。
【００６３】
染色された生体試料の画像を撮影して、色空間で対象物質（粒子）を抽出する方法は、幾つかの方法が知られている。しかし、尿沈渣検査装置では、領域分割を行なう対象が、先に説明したように、変化に富んた性質を有しており、従来法の領域分割方法は、そのまま単純に適用できない。
【００６４】
本発明の発明者の研究によれば、尿沈渣検査装置で得られる画像では、背景領域は、通常、安定に存在することが判明しているので、目的とする対象粒子を検出するために、本発明では、安定に存在する背景領域を先ず識別して、これを基準とする方法を採用した。即ち、背景領域が安定に抽出できれば、背景領域以外の領域に、背景領域と区別して検出すべき対象粒子を表す信号成分が、全て含まれていると考えることができる。
【００６５】
顕微分光光度計による、代表的な沈渣粒子を測定した結果から、染色剤の吸収ピークが５５０ｎｍ近辺にあり、波長領域４００〜５００ｎｍ（青成分）での感度に比較し、波長領域５００〜７００ｎｍ（緑成分、赤成分）での感度が高いことが判明した。この結果から、Ｓ染色で染色された試料の画像の場合には、領域分割を、Ｇ画像、及びＲ画像空間で行なうことが好ましいことが明らかとなった。 図４は、色空間における背景部の抽出を行なう領域分割を模式的に示す。ｘ軸をＧ画像の濃度、ｙ軸をＲ画像の濃度とする。ここで、閾値Ｔ１、Ｔ２は、Ｒ画像の濃度ヒストグラム４０から、閾値Ｔ３、Ｔ４は、Ｇ画像の濃度ヒストグラム４０’から、それぞれ求められる（なお、図４に示す領域分割法と類似する方法が、特開平６−３１４３３８号公報に記載されているが、特開平６−３１４３３８号公報では、先に説明した撮像光学系の歪みによる濃度むらの補正はなされていない）。
【００６６】
具体的には、各々の画像の濃度ヒストグラムを求め、ヒストグラムの最大のピーク値を持つ濃度値、半値幅を与える濃度の値を求める。次いで、これらの値を基にして、対象領域を抽出するに必要な背景領域を識別して取り出す。
【００６７】
以下、各閾値を算出し、背景領域を識別し、識別された目的とする対象粒子のパターンについての特徴量を求める手順について、説明する。
【００６８】
（手順１）：各画像（Ｇ画像、Ｒ画像）の濃度ヒストグラムを作成する。各画像の画素の位置を（ｉ，ｊ）で表わし、Ｇ画像をＧ（ｉ，ｊ）、Ｒ画像をＲ（ｉ，ｊ）と、それぞれ表わす。
【００６９】
（手順２）：各濃度ヒストグラムで、図５に示すように頻度の最大値Ｐmax（＊）をもつ濃度値Ｐ_d（＊）を求める。なお、＊は、Ｒ、Ｇのをいずれかを表わす。
【００７０】
（手順３）：各濃度ヒストグラムで、ピークの半値幅を与える濃度、Ｐ_hL（＊）、Ｐ_hh（＊）を求める。＊は、Ｒ、Ｇのをいずれかを表わす。
【００７１】
（手順４）：（手順１）〜（手順３）で求めた濃度値Ｐ_d（＊）、Ｐ_hL（＊）、Ｐ_hh（＊）を使用して、次の（数５）〜（数８）により、閾値Ｔ１、Ｔ２、Ｔ３、Ｔ４を、それぞれ算出する。Ｒ画像から、
Ｔ１＝Ｐ_d（Ｒ）−｛Ｐ_d（Ｒ）−Ｐ_hL（Ｒ）｝・σ …（数５）
Ｔ２＝Ｐ_d（Ｒ）−｛Ｐ_d（Ｒ）−Ｐ_hh（Ｒ）｝・σ …（数６）
Ｇ画像から、
Ｔ３＝Ｐ_d（Ｇ）−｛Ｐ_d（Ｇ）−Ｐ_hL（Ｇ）｝・σ …（数７）
Ｔ４＝Ｐ_d（Ｇ）−｛Ｐ_d（Ｇ）−Ｐ_hh（Ｇ）｝・σ …（数８）
を得る。ここで、σは、予め定められる、実験的に決められるパラメータであり、σは、通常、３．０〜３．５に設定される。
【００７２】
（手順５）：（手順１）〜（手順４）で算出された閾値Ｔ１〜Ｔ４をもとに、次の（数９）に従い、画像空間で２値画像を作成する。背景領域（ＢＧ）は、
｛Ｔ１≦Ｒ（ｉ，ｊ）＜Ｔ２｝∩｛Ｔ３≦Ｇ（ｉ，ｊ）＜Ｔ４｝ …（数９）
により、得られる。ここで、∩は論理積を表す。
【００７３】
（手順６）：（手順５）で求めた２値画像の修正、整形の処理を行なう。
【００７４】
（手順７）：（手順５）で求めた背景領域以外の、それぞれ互いに独立した領域に、対象粒子が存在する対象領域として、ラベリングを行なう。
【００７５】
（手順８）：（手順７）でラベリングされた対象領域のパターンについて、特徴量を求め、対象領域のパターンを識別する。なお、周知の特徴量を、周知の技術によって求め、個々の対象領域のパターンの識別を行なうが、色調別に抽出した対象パターンを、対象粒子の染色程度や、複数の対象粒子の位置状態（接触や重畳の程度）により、（手順６）で得たパターンを合成すれば、さらに安定な対象パターンを得ることができる。
【００７６】
本実施例での処理の工程を整理すると、以下のようになる。
【００７７】
（１）濃度補正処理：Ｇ画像、及びＲ画像で、光学系の歪みに由来する濃度むらを除去する。
【００７８】
（２）領域分割：Ｇ画像、及びＲ画像で複数の閾値を使用して、背景領域と対象領域とに分割し、背景領域を０、対象領域を１とする２値画像を作成する。
【００７９】
（３）修正処理：対象領域の穴埋め、背景領域のノイズ除去等の２値画像の修正、整形を行なう。
【００８０】
（４）ラベリング：２値画像の連結成分毎にラベリング処理を行い、画像中の複数の対象に番号をつける。
【００８１】
（５）特徴量算出：番号をつけたそれぞれの対象につき、面積、周囲長、平均濃度等の特徴量を求める。
【００８２】
（６）識別：求めた特徴量をもとに、それぞれの対象が、どのような尿沈渣成分であるかを判断し、分類する。
【００８３】
なお、（３）〜（６）では、太め処理、細め処理等のフィルタ処理を含め、周知の従来技術を使用できる。
【００８４】
本実施例を、実際の４サンプルから得た９９画像に適用した結果、９９画像中の９３画像において、対象粒子を領域分割して抽出できた。抽出できなかった６画像は、対象粒子が無染色状態の画像、焦点位置の調整不良による画像等であった。
【００８５】
図６、図７に本実施例を、実際のサンプルに適用した結果（上記（３）の処理済み）を示す。なお、図６、図７では、横（ｘ）軸、縦（ｙ）軸は、それぞれ０〜５１１の５１２画素からなっている。図６は、対象粒子の染色が良好な染色細胞の画像の領域分割結果例、図７は、対象粒子の染色がほとんどされていない細胞の画像の領域分割結果例である。本実施例によれば、図７に示す例のように、対象粒子の染色が十分でない場合、例えば、細胞質濃度と背景濃度に差がなく、完全に抽出できない場合があるが、このような粒子は、以下で説明する、実施例２、実施例３、実施例４による方法により、抽出できる。
【００８６】
（実施例２）
以下では、簡単のため、染色液でほとんど染色されない細胞（粒子）を、無染色細胞（粒子）、また、良好に染色された細胞（粒子）を、染色細胞（粒子）と略記する。なお、淡く染色されるが、染色が不十分で、染色が良好でない細胞（粒子）も、以下では無染色細胞（粒子）と呼ぶことにする。
【００８７】
目的とする対象粒子が、染色剤により良好に染色される染色粒子の場合には、先に説明した実施例１の領域分割方法により、良好な領域分割ができる。
【００８８】
しかし、実際には活性度の高い細胞は染色性が悪く、ほとんど染色されない細胞（無染色細胞）が存在する場合がある。このような無染色細胞は、尿試料中に、高い頻度で存在することが判明している。また、尿試料中には、染色性が良好ではなく、染色が不十分である粒子も存在する。
【００８９】
無染色細胞を含む画像では、対象領域と背景領域との色調が、ほぼ同じであり、対象領域の内部に、背景領域と同程度、又は以下の濃度をもつ部分が多数存在する。無染色細胞が、画像の中に存在する場合、実施例１の領域分割方法では、例えば、先に示したように、図７に示す程度の領域分割しかできず、細胞が存在する領域が同じ値をもつ画素で連結されていない。図７のように、正確な２値画像が得られない場合には、対象領域の特徴量が正確に求められず、分類を誤る可能性が非常に高くなる。沈渣成分の識別精度を向上させるには、染色性が悪く、ほとんど染色されない細胞、その他粒子等の対象を、正確に識別する領域分割が、必要となる。
【００９０】
無染色細胞（染色液でほとんど染色されない細胞）は、ほぼ透明であり、細胞が存在する領域（対象領域）は、背景領域とほぼ同じ色調をしており、対象領域の内部に、背景領域とほぼ同程度の濃度領域を多数含んでいる。まず、無染色細胞を含む画像と、染色細胞を含む画像との差異について説明する。
【００９１】
図８、図９に、それぞれ染色細胞、無染色細胞を含む画像の濃度値の例を示す。
【００９２】
以下に示す図８から図１３における、横軸は、画像のｙ方向での、０〜５１１の５１２画素の位置を示す。
【００９３】
図８に、染色細胞（良好に染色された細胞）を含む画像での、染色細胞を含む領域を通るｙ方向での１ライン（ｘ＝３００）についての濃度値のプロットを示す。図９に、無染色細胞を含む画像での、無染色細胞を含む領域を通るｙ方向での１ライン（ｘ＝３４０）についての濃度値のプロットを示す。
【００９４】
図８、図９において、横軸は、画像中での位置（ｙ座標値）、縦軸は濃度値である。濃度値は、０に近いほど色が濃く、２５５に近いほど色が薄い。図８では、染色細胞の領域の濃度値が、背景領域に比較して、低い値となっている様子が、Ｇ画像に顕著に表れている。この場合、図８に示すように、濃度値に、例えば、〜１８０、〜２２０の値の閾値を設定すると、Ｇ画像において、背景領域と対象領域とを、明確に分割できることが分かる（図８には、参考のために、〜１９０の閾値を同じに図示してある）。図９と、実際の画像（図示せず）との比較から、周囲に比べ、濃度変化の大きい部分（ｙ＝４００の近傍）に、細胞が存在することが判明した。図９から、無染色細胞の領域（対象領域）では、背景の領域に比較し、濃度値の大きな部分、及び小さな部分を、それぞれ多数含み、背景の領域と同程度の濃度値をもつ部分も多数含み、図８の場合のように、濃度値に、単に図９に示すように、同じ閾値を設定するだけでは、正確な領域分割ができないことを示している（図９には、図８と同じ閾値を図示してある）。
【００９５】
さらに、図９から、無染色細胞を含む画像では、背景領域の濃度値の変動は小さいが、無染色細胞の領域（対象領域）の濃度値の変動は大きいことが分かる。
【００９６】
本実施例では、この濃度値の変動に着目して、画像中の濃度値の変化の大きさを計算し、濃度値、及び濃度値の変化の大きさに、それぞれ閾値を設け、領域分割を行ない、無染色細胞（染色液でほとんど染色されない細胞）を含む画像の領域分割を正確に行なう。
【００９７】
まず、濃度値の変化の大きさを表わす指標として、局所的な小領域における、濃度値の分散を考える。画像上の（ｘ，ｙ）における濃度値をＰ（ｘ，ｙ）、局所的分散をｑ（ｘ，ｙ）として、ｑ（ｘ，ｙ）を次の（数１０）で定義する。
【００９８】
ｑ（ｘ，ｙ）＝｛１／（２ｎ＋１）²｝ΣΣ｛Ｐ（ｉ，ｊ）｝²−｛１／（２ｎ＋１）⁴｝｛ΣΣＰ（ｉ，ｊ）｝² …（数１０）
（数１０）において、加算Σは、ｉ＝ｘ−ｎからｉ＝ｘ＋ｎ、及びｊ＝ｙ−ｎからｊ＝ｙ＋ｎについて行なう。
【００９９】
図１０、図１１に、それぞれ染色細胞、無染色細胞を含む画像の分散値の例（ｎ＝２とした時のｑ（ｘ，ｙ）の例）を示す。図１０、図１１は、それぞれ図８、図９と同じ位置（１ライン）における、分散値ｑ（ｘ，ｙ）をプロットしたもので、横軸は画像のｙ座標値、縦軸は（数１０）により計算される分散値である。
【０１００】
また、図１０、図１１において、横軸、ｙ＝４９０近辺の大きなピークは、画像の端部の不連続成分によるもので、画像中の認識対象（細胞）とは無関係である。図１１は、無染色細胞を含む画像の分散値ｑ（ｘ，ｙ）であり、図８と比較すると、背景領域の分散値は０に近く、細胞の存在する領域の分散値は、大きい値となっている。そこで、分散値に、閾値ｑ_thを設定し、ｑ_thより大きい分散値をとる領域を、目的とする対象領域とみなし、無染色細胞を含む画像の領域分割を行なうことにする。図１１には、閾値ｑ_thを、〜２００とする例を示してある。
【０１０１】
図１０は、染色細胞を含む画像の分散値ｑ（ｘ，ｙ）を示し、図８と比較すると、細胞の存在する領域のエッジ部分（横軸、ｙ＝２００、及びｙ＝３００近辺）において、分散値が大きな値を示している。しかし、細胞の内部は、一様な色調で染色されているため、分散値は小さく、分散値に閾値を設定し領域分割を行なうと、細胞のエッジ部分だけが、対象領域として分割されてしまう（図１０には、図１１と同様に、閾値ｑ_thを、〜２００とする例を示してある。）。そこで、分散値に閾値を設定し領域分割を行なうとともに、濃度値に閾値を設定し領域分割を行い、両方の領域分割の結果を重ね合わせることにより、染色細胞と無染色細胞の両方の領域分割を正確に行なうことができる。
【０１０２】
以上で述べた方法では、画像における濃度変動の大きさを検出するために、（数１０）で定義される濃度の局所的分散値を用いたが、装置構成をより単純するため、より簡便に濃度変動の大きさを検出する方法として、次の（数１１）で定義される濃度差分値を用いることにする。（ｘ，ｙ）における濃度値をＰ（ｘ，ｙ）、濃度差分値をｒ（ｘ，ｙ）とする。
【０１０３】
ｒ（ｘ，ｙ）＝ΣＰ（ｘ，ｙ＋ｉ）−ΣＰ（ｘ，ｙ−ｉ） …（数１１）
（数１１）において、加算Σは、ｉ＝１からｉ＝ｎについ行なう。この（数１１）の計算は、加算と減算のみで実行でき、画像の一点で、差分値を計算するために使用する濃度値は、画像上で１次元的に配列する画素の濃度値であるので、簡便な装置構成により、画像信号の転送に同期した高速な処理を実行できる。
【０１０４】
図１２、図１３に、それぞれ染色細胞、無染色細胞を含む画像の濃度差分値（ｎ＝２とした時の濃度差分値）の例を示す。図１２、図１３は、それぞれ図８、図９と同じ位置（１ライン）における濃度差分値をプロットしたものである。分散値の場合と同様、横軸４９０近辺のピーク値は、画像端の不連続部分による。
【０１０５】
図１２は、染色細胞を含む画像の差分値を示しており、図８と比較すると、対象領域のエッジ部分で、差分値は正負に大きく変化する値をとり、対象領域の内部、及び背景領域では、差分値は０に近い値をとっている。
【０１０６】
図１３は、無染色細胞を含む画像の差分値を示し、図９と比較すると分かるように、差分値は背景領域で０に近い値をとり、差分値は、対象（細胞）領域で正負に大きく変化する値をとっている。
【０１０７】
ただし、対象領域の内部でも、差分値が０に近い値をとる部分が、多数含まれているため、差分値に正負の２つの閾値、ｒ_th1、ｒ_th2（ｒ_th1＜０＜ｒ_th2）を設定し、ｒ_th1＜ｒ（ｘ，ｙ）＜ｒ_th2なる領域を、背景領域とする領域分割を行なうと、対象領域の内部で、差分値が０に近い値をとる領域は、背景領域として分割されてしまう。また、この方法を、染色細胞を含む画像に適用すると、細胞のエッジ部分のみを、対象領域として分割することになってしまう。図１２、図１３では、閾値を、ｒ_th1が〜−２０、ｒ_th2が〜２０の場合を、示している。
【０１０８】
ここで、濃度値と濃度差分値との関係を考えてみる。濃度差分値が０をとる点は、濃度値が極大値あるいは極小値をとっている点であり、このような点における濃度値は、背景領域の平均の濃度値より比較的大きな値、あるいは小さい値をとる可能性が高い。つまり、濃度差分値に閾値を設定して領域分割を行なった結果、正確に領域分割できない部分（背景領域として分割されてしまう対象領域の内部の部分）は、濃度値に閾値を設定し、領域分割を行えば対象領域として分割できる。そこで、差分値に閾値を設定し領域分割を行なうとともに、濃度値に閾値を設定し領域分割を行ない、両方の領域分割の結果を重ね合わせること（論理ＯＲをとる）により、無染色細胞を含む画像、染色細胞を含む画像の両方を、正確に領域分割することができる。
【０１０９】
以上説明した方法を、実際にフロー方式で撮像した画像に適用した結果について説明する。最も無染色となりやすい偏平上皮細胞を含む画像の５１例を調べたところ、無染色粒子を含む画像が、８例出現した。実施例１の方法では、正確な領域分割が困難であった無染色粒子を含む８例の画像に、以上説明した方法を適用した結果、染色粒子を含む画像の場合に得られる処理結果と、同等の結果を得ることができた。以下では、これらの画像のうち、典型的な染色粒子を含む画像と、無染色粒子を含む画像とを、それぞれ一例ずつとって説明する。
【０１１０】
図１４、図１５に、濃度の局所的分散を用いた２値化による方法を用いて、それぞれ染色細胞、無染色細胞を含む画像を領域分割を行った結果（２値画像（領域分割された２値画像に修正処理を行なった画像））を示す。
【０１１１】
なお、図１４から図１７において、横（ｘ）軸、及び縦（ｙ）軸は、それぞれ０〜５１１の５１２画素からなっている。
【０１１２】
Ｇ画像の濃度値を使用して、（数１０）で、ｎ＝２として計算される分散値と、Ｇ画像の濃度値を、領域分割の処理で用いた。具体的には、（ｘ，ｙ）における濃度値をＰ（ｘ，ｙ）、分散値をｑ（ｘ，ｙ）、２値化結果をｓ（ｘ，ｙ）とすると、
Ｐ_th1＜Ｐ（ｘ，ｙ）＜Ｐ_th2 ａｎｄ ｑ（ｘ，ｙ）＜ｑ_th：
ｓ（ｘ，ｙ）＝０（背景：白）
ｏｔｈｅｒｗｉｓｅ：ｓ（ｘ，ｙ）＝１（対象領域：黒） …（数１２）
として、ｓ（ｘ，ｙ）で表される２値画像を求めた後、修正処理（背景領域のノイズ除去処理、対象領域の穴埋め処理）として、対象領域の太め処理を１回行い、細め処理を２回行った後、さらに太め処理を１回行った（これらの各処理は周知技術である）。
【０１１３】
濃度値に設定する閾値Ｐ_th1、Ｐ_th2は、濃度ヒストグラムの最大値を与える濃度値Ｐ_m、最大値の１／２の値を与える濃度値、Ｐ_s1、Ｐ_s2（Ｐ_s1＜Ｐ_m＜Ｐ_s2）を求め、
Ｐ_th1＝Ｐ_m−３．５（Ｐ_m−Ｐ_s1）
Ｐ_th2＝Ｐ_m＋３．５（Ｐ_s2-Ｐ_m） …（数１３）
とした。また、分散値に設定する閾値はｑ_th＝３０とした。
【０１１４】
図１５より、明らかに、無染色細胞を含む画像の領域分割の結果は、図７に示す、実施例１による結果に比較し、大きく改善されている。
【０１１５】
また、染色細胞を含む画像（図１４）において、細菌が、細胞の領域の右側の２箇所（５０、５１）に、無染色細胞を含む画像（図１５）において、細菌が、細胞の領域の右側の１箇所（５２）に、それぞれ検出されている。これらの細菌は、実施例１の方法による結果（図６、図７）では、検出できなかった細菌である。
【０１１６】
染色細胞を含む画像（図１４）の細胞の領域の左側の２箇所（６０、６１）に、無染色細胞を含む画像（図１５）の細胞の領域の左側に１箇所（６２）に、細菌より小さな点があるが、これは背景領域のノイズ成分と思われる。
【０１１７】
図１６、図１７に、濃度差分値を用いた２値化による方法を用いて、それぞれ染色細胞、無染色細胞を含む画像を領域分割を行った結果（２値画像（領域分割された２値画像に修正処理を行なった画像））を示す。Ｇ画像の濃度値を使用して、（数１１）で、ｎ＝２として計算される差分値、及びＧ画像の濃度値を、領域分割の処理で用いた。具体的には、（ｘ，ｙ）における濃度値をＰ（ｘ，ｙ）、差分値をｒ（ｘ，ｙ）、２値化結果をｓ（ｘ，ｙ）とすると、
Ｐ_th1＜Ｐ（ｘ，ｙ）＜Ｐ_th ａｎｄ ｒ_th1＜ｒ（ｘ，ｙ）＜ｒ_th2：
ｓ（ｘ，ｙ）＝０（背景：白）
ｏｔｈｅｒｗｉｓｅ：ｓ（ｘ，ｙ）＝１（対象領域：黒） …（数１４）
として、ｓ（ｘ，ｙ）で表される２値画像を求めた後、修正処理を行った。
【０１１８】
修正処理、及び濃度閾値Ｐ_th1、Ｐ_th2の設定は、上記の濃度の局所的分散を用いた２値化による方法を用いた領域分割の場合と、同じ方法を用いた。また、差分値に設定する閾値は、│ｒ_th1│、│ｒ_th2│は１５〜２５程度に設定すればよく、ｒ_th1＝−２０、ｒ_th2＝２０とした。
【０１１９】
図１７より、明らかに、無染色細胞を含む画像の領域分割の結果は、図７に示す、実施例１による結果に比較し、大きく改善されている。
【０１２０】
また、染色細胞を含む画像（図１６）において、細菌が、細胞の領域の右側の２箇所（５０、５１）に、無染色細胞を含む画像（図１７）において、細菌が、細胞の領域の右側の１箇所（５２）に、それぞれ検出されている。これらの細菌は、実施例１の方法のよる結果（図６、図７）では、検出できなかった細菌である。
【０１２１】
図７、図１５、図１７を比較すると、実施例１の領域分割法に比較し、上記の濃度の局所的分散、差分値を用いた２値化による方法を用いた２種類の領域分割法は、無染色細胞の領域分割に効果的であることが分かる。また、２種類の方法のうち、濃度の分散値を用いる方法の方が、無染色細胞の領域分割に、より効果的である。しかし、濃度の差分値を用いる方法の方が、装置が単純に構成できる。そこで、差分値を用いた領域分割法の詳細を説明する。
【０１２２】
図１６、図１７に示す結果は、差分値の計算（数１１）において、ｎ＝２とした時の結果である。ある画素の差分値を計算する際に用いる近傍画素の個数は２ｎ＋１で表され、この個数を、ここではマスクサイズと呼ぶことにする。例えば、ｎ＝２の時、ある画素の差分値の計算には、その画素を中心とし、ｙ方向に前後２画素、合計５画素が必要であり、マスクサイズ＝５である。マスクサイズ＝１の場合は、差分処理を行なわないことを意味する。
【０１２３】
図１８に、マスクサイズの変化による領域分割の結果を示す。図１８では、５１２×５１２画素からなる画像を示している。マスクサイズが、領域分割に与える影響を調べるために、ｎの値を変化させて、領域分割を行った。図１８の例は、図７、図１５、図１７とそれぞれ同じ無染色細胞を含む画像の領域分割の結果である。図８における、マスクサイズ＝５の結果は、図１７の結果に相当する。図１８の例では、マスクサイズ＝９による結果が、もっとも良好な領域分割の結果を与えている（しかし、細胞が存在する領域の内部の一部分に、誤って識別された背景を示す白い領域が混在している）。
【０１２４】
また、図１５、図１７の結果と同様に、図１８においても、細胞５２が、検出されている。なお、マスクサイズ＝１による差分処理を行なわない場合の結果は、図７に示す、実施例１による結果に類似した結果である。
【０１２５】
次に、このマスクサイズの大きさの影響を周波数軸上で考えてみる。
【０１２６】
図１９に、点線で、背景領域のｙ方向に１ライン（５１２画素）のＧ画像の信号成分（濃度値）を、離散フーリエ変換した結果を、実線で、無染色細胞の領域を通るｙ方向の１ライン（５１２画素）のＧ画像の信号成分（濃度値）を、離散フーリエ変換した結果を、それぞれ示す。横軸は、画像の１辺（５１２画素）を単位長として表現した空間周波数、縦軸は、濃度値の最大値を０．１に正規化したときの信号のパワーを表わしている。
【０１２７】
図１９から、背景領域の信号はノイズ成分であるため、直流成分を除き、ほぼ全域に同等のパワーをもち分布していること、無染色細胞の領域を通過するラインの画像信号は、背景領域に比較し、周波数１〜１５０の信号成分が強い、ということが分かる。
【０１２８】
無染色細胞の領域を強調するために用いた（数１１）を、ディジタルフィルタとみなすと、伝達関数はｚ変換を用いて、
Ｈ（ｚ）＝Σ｛ｚ↑（ｋ）−ｚ↑（−ｋ）｝ …（数１５）
と表される。ここでｚ↑（ｙ）は、ｚのべき（ｙ）乗を示す（ｙ＝ｋ、ｙ＝−ｋ）。（数１５）において、加算Σは、ｋ＝１からｋ＝ｎについて行なう。また、この時の周波数応答は、

と表される。Ωは、画像上の２画素を単位長とする空間周波数を表わし、ｊは虚数単位である。
【０１２９】
図２０に、ｎ＝１、２、３の場合の周波数振幅特性｜Ｈ（Ω）｜）を示す。
【０１３０】
図２０の横軸は、画像の１辺の長さ（５１２画素）を単位長として表現した空間周波数を表し、縦軸は、フィルタの応答をマスクサイズを用いて正規化した値を表している。また、位相特性はπ／２で一定であり、このフィルタの作用は、無染色細胞を含む画像に特有の周波数成分を強調し、背景から分離する作用と、位相を９０度ずらすことにより、濃度で２値化したとき０になる部分と、差分値を用いて２値化したとき０になる部分とをずらす作用と、を有している。
【０１３１】
図２０から、ｎ＝２（マスクサイズ５）、ｎ＝３（マスクサイズ７）の場合には、無染色細胞を含む画像に特有の中周波領域〜低周波領域の信号成分が強められるが、マスクサイズが過大になると、図２０には、図示していないが、ｎ＝５の場合のように、ｎの値を大きくしていくに従い、無染色細胞を含む画像に特有の信号成分のうち、低周波領域が弱められてしまう。このため、ｎの値は、２〜４（マスクサイズ５〜９）程度が適当である。
【０１３２】
図２１（Ａ）、図２１（Ｂ）に、染色された赤血球の画像を用いて、対象領域として分割される領域の形状の、マスクサイズによる影響を、調べた結果を示す。図２１（Ａ）、図２１（Ｂ）において、長さの単位は、画素数である。
【０１３３】
分類すべき対象のうち、赤血球は、細菌に次いで小さく、その形状がほぼ円形であるので、赤血球像を、マスクサイズを変化させて領域分割し、抽出された対象領域の形状の変化を評価することにより、画像処理過程における劣化（対象粒子がもつ真の形状と、領域分割された対象領域の形状との差）を評価できる。
【０１３４】
図２１（Ａ）は、種々のマスクサイズ（図中に数字で示す）を用いた場合について得た、対象領域（赤血球）の面積、周囲長のプロットである。マスクサイズが大きくなるにつれ、面積、周囲長ともに大きくなり、対象領域が広がっていく。
【０１３５】
また、図２１（Ｂ）は、領域分割された赤血球を、それぞれｘ軸、ｙ軸へ投影した時に得る、ｘ、及びｙ方向の投影長のプロットである。赤血球はほぼ円形をしているので、差分処理を用いない場合（マスクサイズ＝１）には、ｘ軸、ｙ軸への投影長が、ほぼ同じ値である。しかし、差分処理を行なうと、ｘ軸方向への投影長に比べ、差分処理の方向であるｙ軸方向の投影長が、より大きくなり、対象領域が、縦長の楕円形になっている。これは、図２０から分かるように、差分処理により、高周波成分が抑制され、対象のエッジが広がるため、処理方向へ伸張されるためである。
【０１３６】
また、マスクサイズが大きくなる程この傾向は強くなる。このように、無染色細胞を含む画像の２値化における改善、染色細胞を含む画像の２値画像の劣化、という２点を考慮して、マスクサイズを決定する必要がある。
【０１３７】
無染色細胞を含む画像の２値化の改善は、より正確な特徴パラメータの算出を可能にし、後段でのパターン認識の認識率の向上につながるが、染色細胞を含む画像の２値画像の劣化は、特徴パラメータを不正確にし、特にサイズの小さな対象の分類精度に影響する。また、２値化の際に用いる閾値によって、得られる２値画像は変化する。一般に、閾値の絶対値が、小さいほど無染色細胞の領域分割は良好に実行できるが、背景領域のノイズ成分を拾いやすくなる。この点に注意して最適な閾値を設定する必要がある。
【０１３８】
図２２に、本実施例の粒子画像の領域分割の手順を説明する処理のフローチャートを示す。先に説明したように、処理の対象となる粒子画像は、画像メモリに記憶されている。本実施例で、使用する画像はＧ画像であり、各画素の位置（ｘ，ｙ）における濃度値Ｐ（ｘ，ｙ）で表わすものとする。
【０１３９】
以下、図２２に示す、処理のステップに従って、処理過程を説明する。
【０１４０】
（ステップＳ１）： まず、画素（ｘ，ｙ）における濃度値Ｐ（ｘ，ｙ）を画像メモリより読み出す。
【０１４１】
（ステップＳ２）： 濃度値Ｐ（ｘ，ｙ）を、以下に示す２値画像を表す変数ｕ（ｘ，ｙ）に変換する。
【０１４２】
Ｔ２’＜Ｐ（ｘ，ｙ）＜Ｔ１’の時：ｕ（ｘ，ｙ）＝０ …（数１７）
ｏｔｈｅｒｗｉｓｅ ：ｕ（ｘ，ｙ）＝１ …（数１８）
ここで、Ｔ１’、Ｔ２’は閾値を表し、例えば、次のようにしてＴ１’、Ｔ２’を定める。まず、画像メモリに記憶された画像の濃度値Ｐ（ｘ，ｙ）を読み出し、濃度値のヒストグラムを作成し、ヒストグラムの形状から濃度閾値Ｔ１’、Ｔ２’を決定する。通常、画像上で背景部分の面積が最も大きいので、背景の平均濃度がヒストグラム上での最大値をとる。図５に示すように、このヒストグラムの最大値を与える濃度値Ｐ_dを検出し、この最大値をＰmaxとする。次に、Ｐmax／２の値を与える濃度値を検出し、Ｐ_hL、Ｐ_hh（Ｐ_hL＜Ｐ_d＜Ｐ_hh）とする。
【０１４３】
適当な定数ｋを定め、閾値Ｔ１’、Ｔ２’を、（数１９）、（２０）から、決定する。
【０１４４】
Ｔ１’＝Ｐ_d＋ｋ（Ｐ_hL−Ｐ_d） …（数１９）
Ｔ２’＝Ｐ_d＋ｋ（Ｐ_hh−Ｐ_d） …（数２０）
定数ｋは、あらかじめ実験的に求めておく。
【０１４５】
（数１７）に示すように、Ｔ２’＜Ｐ（ｘ，ｙ）＜Ｔ１’の時、画素（ｘ，ｙ）は背景領域にあるとみなし（（数１７））、Ｔ２’≧Ｐ（ｘ，ｙ）、又はＰ（ｘ，ｙ）≧Ｔ１’の時、画素（ｘ，ｙ）は対象領域（背景領域以外の領域であり、背景領域と識別しようとする粒子が存在する領域をさす）にあるとみなす（（数１８））。
【０１４６】
（ステップＳ３）： 画素（ｘ，ｙ）の近傍の複数の画素の濃度値を、メモリより読み出す。
【０１４７】
（ステップＳ４）： 濃度値の変化の大きさ（画素（ｘ，ｙ）における濃度値の変化の大きさをｑ（ｘ，ｙ）とする。）を、下記（ａ）〜（ｄ）の何れかの方法を用いて、計算する。
【０１４８】
（ａ）；（数２１）、又は（数２２）に示すように、ステップＳ３で読み出した、画素（ｘ，ｙ）の近傍の複数の画素の濃度値を、画像上で近接する２つの領域に属する画素に分け、それぞれの領域に属する画素の濃度値の和を求め、次に、得られた２つの和の値の差（（数２１））、又はその絶対値（（数２２））を求める。
【０１４９】
ｑ（ｘ，ｙ）＝Σ｛Ｐ（ｘ，ｙ＋ｉ）−Ｐ（ｘ，ｙ−ｉ）｝ …（数２１）
ｑ（ｘ，ｙ）＝│Σ（Ｐ（ｘ，ｙ＋ｉ）−Ｐ（ｘ，ｙ−ｉ））│ …（数２２）
例えば、ステップＳ３で読み出した画素（ｘ，ｙ）の近傍の４個の画素の濃度値、Ｐ（ｘ，ｙ−２）、Ｐ（ｘ，ｙ−１）、Ｐ（ｘ，ｙ＋１）、Ｐ（ｘ，ｙ＋２）から、（数２１）、又は（数２２）によって、ｑ（ｘ，ｙ）を計算する。(数２１)、（数２２）において、加算Σはｉ＝１からｉ＝ｎについて行なう。
【０１５０】
（ｂ）；ステップＳ３で読み出した、画素（ｘ，ｙ）の近傍の複数の画素の濃度値の分散（（数２３））、又は標準偏差（（数２４））を求め、濃度値の変化の大きさｑ（ｘ，ｙ）とする。
【０１５１】
ｑ（ｘ，ｙ）＝｛１／（２ｎ＋１）²｝ΣΣ｛Ｐ（ｘ＋ｉ，ｙ＋ｊ）｝²−｛１／（２ｎ＋１）²｝²｛ΣΣＰ（ｘ＋ｉ，ｙ＋ｊ）｝² …（数２３）
ｑ（ｘ，ｙ）＝√〈｛１／（２ｎ＋１）²｝ΣΣ｛Ｐ（ｘ＋ｉ，ｙ＋ｊ）｝² −｛１／（２ｎ＋１）²｝²｛ΣΣＰ（ｘ＋ｉ，ｙ＋ｊ）｝²〉 …（数２４）
なお、√〈 〉は、〈 〉内の平方根を示す。（数２３）、（数２４）において加算Σは、それぞれｉ＝−ｎからｉ＝ｎ、ｊ＝−ｎからｊ＝ｎについて行なう。
【０１５２】
例えば、ステップＳ３において、ｘ−１≦ｉ≦ｘ＋１、ｙ−１≦ｊ≦ｙ＋１を満たす９個の画素の濃度値Ｐ（ｉ，ｊ）を読み出し、（数２３）、又は（数２４）によって、ｑ（ｘ，ｙ）を計算する。
【０１５３】
（ｃ）；ステップＳ３で得た画素（ｘ，ｙ）の近傍の複数の画素の濃度値と、それぞれの画素の幾何学的配置に対応した重み値ｗと、の積をとり、得られた積の総和を濃度値の変化の大きさｑ（ｘ，ｙ）（（数２５））とする。
【０１５４】
ｑ（ｘ，ｙ）＝｛ΣΣｗ（ｉ，ｊ）Ｐ（ｘ＋ｉ，ｙ＋ｊ）｝ …（数２５）
（数２５）において、加算Σはそれぞれｉ＝−ｎからｉ＝ｎ、及びｊ＝−ｎからｊ＝ｎについて行なう。例えば、９個の重み値ｗ（ｋ，ｍ）（ｋ，ｍ＝−１、０、１）をあらかじめ設定し、ステップＳ３において、９個の画素の濃度値Ｐ（ｉ，ｊ）（ｘ−１≦ｉ≦ｘ＋１、ｙ−１≦ｊ≦ｙ＋１）を読み出し、（数２５）によってｑ（ｘ，ｙ）を計算する。重み値は、例えば、ｗ（０，０）＝４、ｗ（−１，０）＝ｗ（１，０）＝ｗ（０，−１）＝ｗ（０，１）＝１、ｗ（−１，−１）＝ｗ（−１，１）＝ｗ（１，−１）＝ｗ（１，１）＝０とする。
【０１５５】
（ｄ）；上記（ａ）〜（ｃ）の方法のうち、何れかの方法を用いて濃度値の変化の大きさｑ（ｘ，ｙ）を求めた後、得られた濃度値の変化の大きさを濃度値Ｐ（ｘ，ｙ）とみなし、
Ｐ（ｘ，ｙ）＝ｑ（ｘ，ｙ） …（数２６）
とおいて、上記（ａ）〜（ｃ）の何れかの方法により、濃度値の変化の大きさｑ（ｘ，ｙ）を計算してもよい。
【０１５６】
例えば、上記（ａ）の方法により、すべての画素（ｘ，ｙ）について、ｑ（ｘ，ｙ）の値を計算した後、（数２６）を適用して、次いで、上記（ｂ）の方法を用いて、再度ｑ（ｘ，ｙ）を計算する。
【０１５７】
（ステップＳ５）： 濃度値の変化の大きさｑ（ｘ，ｙ）を、以下に示す２値画像を表す変数ｖ（ｘ，ｙ）に変換する。
【０１５８】
ｑ（ｘ，ｙ）が正値のみを取る場合は、閾値Ｔ３’を使用して、
ｑ（ｘ，ｙ）＜Ｔ３’の時：ｖ（ｘ，ｙ）＝０ …（数２７）
ｑ（ｘ，ｙ）≧Ｔ３’の時：ｖ（ｘ，ｙ）＝１ …（数２８）
とし、ｑ（ｘ，ｙ）＜Ｔ３’の時、画素（ｘ，ｙ）は背景領域にあるとみなし（（数２７））、ｑ（ｘ，ｙ）≧Ｔ３’の時、画素（ｘ，ｙ）は対象領域にあるとみなす（（数２８））。なお、上記の閾値Ｔ３’は、予め実験的に求めておくことができる。
【０１５９】
ｑ（ｘ，ｙ）が負値も取る場合（ステップＳ５’）は、閾値Ｔ３”、閾値Ｔ４’を使用して、
Ｔ４’＜ｑ（ｘ，ｙ）＜Ｔ３”の時：ｖ（ｘ，ｙ）＝０ …（数２９）
ｏｔｈｅｒｗｉｓｅ ：ｖ（ｘ，ｙ）＝１ …（数３０）
とし、Ｔ４’＜ｑ（ｘ，ｙ）＜Ｔ３”の時、画素（ｘ，ｙ）は背景領域にあるとみなし（（数２９））、Ｔ４’≧ｑ（ｘ，ｙ）、又はｑ（ｘ，ｙ）≧Ｔ３”の時、画素（ｘ，ｙ）は対象領域にあるとみなす（（数３０））。なお、上記の閾値Ｔ３”、閾値Ｔ４’は、予め実験的に求めておくことができる。
【０１６０】
（ステップＳ４’）でのＳ５、Ｓ５’の選択は、（ステップＳ４）において、どのようにｑ（ｘ，ｙ）が定義されるかにより、自動的に定まる。
【０１６１】
（ステップＳ６）： （ステップＳ２）で得られたｕ、と（ステップＳ５）、又は（ステップＳ５’）で得られたｖの論理和を、画素（ｘ，ｙ）における２値化の結果とする。
【０１６２】
（ステップＳ７、Ｓ８）： 画素（ｘ，ｙ）が処理すべき最終の画素の点であれば、処理をステップＳ９に移し、そうでなければ画素（ｘ，ｙ）を、次の画素の点に移動して、ステップＳ１〜Ｓ６の処理を繰り返す（ステップＳ７、Ｓ８）。
【０１６３】
（ステップＳ９）： 以上の処理により得た、対象領域を１、背景領域を０とする２値画像に対して、対象領域の穴埋め、背景領域のノイズ除去等の処理を行なう修正処理を行なう。
【０１６４】
（ステップＳ１０）： ２値画像の連結成分毎に、周知技術によるラべリング処理を行ない、画像中の複数の対象に番号を付ける。
【０１６５】
（ステップＳ１１）： 番号が付けられた各対象について、面積、周囲長、平均濃度、投影長等の、周知技術による特徴量を求める。
【０１６６】
（ステップＳ１２）： 求められた特徴量を使用して、各対象を、尿沈渣成分のいずれかの種類に分類識別する。
【０１６７】
図２３に、本実施例を実施するための装置構成を示す。液体中の粒子はパルスランプ１０１により照射され、光学拡大系（図示せず）によって拡大された後、テレビカメラ１０２により撮影され、アナログ電気信号に変換される。テレビカメラ１０２の出力は、Ａ／Ｄ変換器１０３によりディジタル信号に変換された後、データバス１１０を介して、画像メモリ１０４に記憶される。閾値計算部１０５では、データバス１１０を介して、画像メモリ１０４に記憶された画像の濃度値を読み出し、濃度値のヒストグラムを作成し、（数１９）、（数２０）に従って、ヒストグラム形状より領域分割部１０６で用いる濃度閾値Ｔ１’、Ｔ２’を決定する。
【０１６８】
領域分割部１０６は、データバス１１０を介して、画像メモリ１０４より得た濃度値と、閾値計算部１０５より得た閾値Ｔ１’、Ｔ２’と、を用いて領域分割を行ない、対象領域を１、背景領域を０とする２値化を行った後、データバス１１０を介して、２値化の結果を、画像メモリ１０７に記憶する。以上の処理の制御は、全て計算機１０８が、ＣＰＵバス１０９を介して行う。
【０１６９】
図２４は、図２３における領域分割部１０６の構成を示す。入力部２００は、図２３のデータバス１１０に接続しており、データバス１１０を介したデータの入力を制御する。濃度２値化回路３００では、まず、入力部２００を介して、閾値計算部１０５（図２３）より、閾値Ｔ１’、Ｔ２’を読み出し、閾値メモリ３１０に、Ｔ１’、Ｔ２’の値を記憶する。次に、入力部２００を介して、画像メモリ１０４から、画像上の一点の画素の濃度値を読み出し、比較器３２０で、この濃度値と閾値Ｔ１’、Ｔ２’とを比較する。画像上の一点を（ｘ，ｙ）、その点における濃度値をＰ（ｘ，ｙ）とすると、比較器３２０では、濃度値Ｐ（ｘ，ｙ）と閾値Ｔ１’、Ｔ２’とを比較し、Ｔ２’＜Ｐ（ｘ，ｙ）＜Ｔ１’であれば、その画素は背景領域であるとみなして、０を出力し、それ以外の時には、対象領域であるとみなして、１を出力する。
【０１７０】
また、上記の濃度の２値化と並行して、濃度変化２値化回路４００が働く。濃度変化２値化回路４００では、まず入力部２００を介して、画像メモリ１０４（図２３）から、画像上（ｘ，ｙ）の複数の近傍画素の濃度値を読み出し、メモリ４１０に記憶する。濃度変化計算部４２０では、メモリ４１０に保持された（ｘ，ｙ）の近傍画素の濃度値を利用して、下記（ａ）〜（ｄ）の何れかの方法を用いて濃度値の変化の大きさを計算する。（ｘ，ｙ）における濃度値の変化の大きさをｑ（ｘ，ｙ）とする。
【０１７１】
（ａ）メモリ４１０に記憶されている画素の濃度値を、画像上で近接する２つの領域に属する画素に分け、それぞれの領域に属する画素の濃度値の和を求める。次に、得られた２つの和の値の差、又はその絶対値を求め、濃度値の変化の大きさをｑ（ｘ，ｙ）とする。例えば、４個の画素の濃度値Ｐ（ｘ，ｙ−２）、Ｐ（ｘ，ｙ−１）、Ｐ（ｘ，ｙ＋１）、Ｐ（ｘ，ｙ＋２）を、メモリ４１０に記憶しておき、（数３１）、又は（数３２）によって、ｑ（ｘ，ｙ）を計算する。
【０１７２】
ｑ（ｘ，ｙ）＝ΣＰ（ｘ，ｙ＋ｉ）−ΣＰ（ｘ，ｙ−ｉ） …（数３１）
ｑ（ｘ，ｙ）＝│ΣＰ（ｘ，ｙ＋ｉ）−ΣＰ（ｘ，ｙ−ｉ）│ …（数３２）
（数３１）、（数３２）において、加算Σはそれぞれｉ＝１からｉ＝２について行なう。図２５に、この場合の濃度変化２値化回路４００の具体的な構成例を示す。メモリ４１０は、１個の濃度値をそれぞれ記憶する４個のメモリ４１１〜４１４から構成される。メモリ４１１〜４１４に、濃度値Ｐ（ｘ，ｙ−２）、Ｐ（ｘ，ｙ−１）、Ｐ（ｘ，ｙ＋１）、Ｐ（ｘ，ｙ＋２）を、それぞれ記憶する。
【０１７３】
濃度変化計算部４２０は、２個の加算器４２１、４２２と１個の減算器４２３より構成され、（数３１）で示される計算を行う。加算器４２１の出力は、（数３１）の右辺の第２項で計算される値を出力し、加算器４２２は、（数３１）の右辺の第１項で計算される値を出力する。減算器４２３の出力は、（数３１）の右辺で計算される値を出力する。減算器４２３に、絶対値を出力する機能を持たせた場合、（数３２）で示される計算をできる。
【０１７４】
（ｂ）メモリ４１０に記憶された濃度値の分散、又は標準偏差を求め、濃度値の変化の大きさｑ（ｘ，ｙ）とする。例えばｘ−１≦ｉ≦ｘ＋１、ｙ−１≦ｊ≦ｙ＋１を満たす、９個の画素の濃度値Ｐ（ｉ，ｊ）をメモリ４１０に記憶しておき、次の（数３３）、又は（数３４）によってｑ（ｘ，ｙ）を計算する。
【０１７５】
ｑ（ｘ，ｙ）＝｛１／９｝ΣΣ｛Ｐ（ｘ＋ｉ，ｙ＋ｊ）｝²−｛１／８１｝｛ΣΣＰ（ｘ＋ｉ，ｙ＋ｊ）｝² …（数３３）
ｑ（ｘ，ｙ）＝√〈｛１／９｝ΣΣ｛Ｐ（ｘ＋ｉ，ｙ＋ｊ）｝²−｛１／８１｝｛ΣΣＰ（ｘ＋ｉ，ｙ＋ｊ）｝²〉 …（数３４）
なお、√〈 〉は、〈 〉内の平方根を示す。（数３３）、（数３４）において、加算Σはそれぞれ、ｉ＝−１からｉ＝１、及びｊ＝−１からｊ＝１について行なう。
【０１７６】
図２６に、この場合のメモリ４１０と濃度変化計算部４２０の具体的な回路構成例を示す。メモリ４１０は、１個の濃度値をそれぞれ記憶する９個のメモリ８１１〜８１９から構成される。メモリ８１１〜８１９には、ｘ−１≦ｉ≦ｘ＋１、ｙ−１≦ｊ≦ｙ＋１を満たす、９個の画素の濃度値Ｐ（ｉ，ｊ）を記憶させる。
【０１７７】
濃度変化計算部４２０は、９個の２乗演算器からなる２乗演算部８２１、加算器８２２、８２４、除算器８２３、８２５、２乗演算器８２６、減算器８２７から構成される。メモリ８１１〜８１９から読み出された値は、それぞれ２つに分けられ、一方は、２乗演算部８２１に入力され、他方は、加算器８２４に入力される。加算器８２２で２乗演算部８２１の９個の２乗演算器の出力の和を計算し、除算器８２３で、加算器８２２の出力値を９で割る。除算器８２３は、（数３３）の右辺の第１項で計算される値を出力する。
【０１７８】
加算器８２４は、メモリ８１１〜８１９から読み出された９個の濃度値の和を計算し、除算器８２５で加算器８２４の出力値を９で割る。２乗演算器８２６は、（数３３）の右辺の第２項で計算される値を出力する。減算器８２７では、除算器８２３の出力と２乗演算器８２６の出力の差を計算する。減算器８２７は、（数３３）の右辺で計算される値を出力する。減算器８２７の後に平方根を計算する回路を設けると、（数３４）の右辺で計算される値を出力できる。
【０１７９】
（ｃ）メモリ４１０に記憶された画素の濃度値と、画素の画像上の幾何学的位置に対応した重み値ｗとの積をとり、得られた積の総和を濃度値の変化の大きさｑ（ｘ，ｙ）とする。例えば、９個の重み値ｗ（ｋ，ｍ）（ｋ，ｍ＝−１、０、１）を予め設定し、メモリ４１０には、９個の画素の濃度値Ｐ（ｉ，ｊ）（ｘ−１≦ｉ≦ｘ＋１、ｙ−１≦ｊ≦ｙ＋１）を保持しておき、次の（数３５）によって、ｑ（ｘ，ｙ）を計算する。
【０１８０】
ｑ（ｘ，ｙ）＝｛ΣΣｗ（ｉ，ｊ）Ｐ（ｘ＋ｉ，ｙ＋ｊ）｝ …（３５）
重み値は、例えば、ｗ（０，０）＝−４、ｗ（−１，０）＝ｗ（１，０）＝ｗ（０，−１）＝ｗ（０，１）＝１、ｗ（−１，−１）＝ｗ（−１，１）＝ｗ（１，−１）＝ｗ（１，１）＝０とする。（数３５）において、加算Σは、それぞれ、ｉ＝−１からｉ＝１、及びｊ＝−１からｊ＝１について行なう。
【０１８１】
図２７に、この場合のメモリ４１０と濃度変化計算部４２０の具体的な回路構成例を示す。メモリ４１０は、１個の濃度値をそれぞれ記憶する９個のメモリ９３１〜９３９から構成される。メモリ９３１〜９３９には、濃度値Ｐ（ｘ−１，ｙ−１）、Ｐ（ｘ，ｙ−１）、Ｐ（ｘ＋１，ｙ−１）、Ｐ（ｘ−１，ｙ）、Ｐ（ｘ，ｙ）、Ｐ（ｘ＋１，ｙ）、Ｐ（ｘ−１，ｙ＋１）、Ｐ（ｘ，ｙ＋１）、Ｐ（ｘ＋１，ｙ＋１）が、それぞれ記憶される。重み値メモリ９４１〜９４９には、重み値ｗ（−１，−１）、ｗ（０，−１）、ｗ（１，−１）、ｗ（−１，０）、ｗ（０，０）、ｗ（１，０）、ｗ（−１，１）、ｗ（０，１）、ｗ（１，１）を、それぞれ予め設定しておく。乗算器９５１〜９５９により、濃度値と重み値の積を計算し、得られた９個の積の和を加算器９６１により求める。
【０１８２】
（ｄ）上記（ａ）〜（ｃ）の方法のうち、何れかの方法を用いて、濃度値の変化の大きさを求めた後、得られた濃度値の変化の大きさを、濃度値とみなし、上記（ａ）〜（ｃ）の方法のうち、何れかの方法を用いて再度濃度値の変化の大きさを計算し、得られた値をｑ（ｘ，ｙ）としてもよい。
【０１８３】
例えば、上記（ｃ）の方法ですべての画素（例えば、５１２×５１２画素）について、ｑ（ｘ，ｙ）の値を計算した後、このｑ（ｘ，ｙ）を用いて、
Ｐ（ｘ，ｙ）＝ｑ（ｘ，ｙ） …（３６）
とおいて、更に上記（ｂ）の方法を用いて、再度ｑ（ｘ，ｙ）を計算する。この場合、メモリ４１０と濃度変化計算部４２０は、図２８で示される構成の濃度変化計算部の回路に置き換えられる。
【０１８４】
まず、荷重和計算回路９００によって、全ての画素について（数３５）の右辺の値を計算し、得られた値を画像メモリ１０００に記憶する。次に、画像メモリ１０００に記憶された値を、濃度値として分散計算回路８００により、（数３３）の右辺の値を計算し出力する。荷重和計算回路９００、分散計算回路８００の構成の構成は、それぞれ図２７、図２６に示される構成である。
【０１８５】
図２４の比較部４３０は、図２５に示すように、閾値メモリ４３１と比較器４３２とからなる。閾値メモリ４３１には、予め閾値Ｔ３’、又はＴ３”、Ｔ４’が記憶されており、濃度変化の大きさｑ（ｘ，ｙ）が正値のみを取る場合、ｑ（ｘ，ｙ）＜Ｔ３’であれば、その画素は背景領域であるとみなし、０を出力し、それ以外の時には対象領域であるとみなし、１を出力する。
【０１８６】
また、ｑ（ｘ，ｙ）が負値も取る場合には、Ｔ４’＜ｑ（ｘ，ｙ）＜Ｔ３”ならば０、それ以外の時には１を出力する。閾値Ｔ３’、Ｔ３”、Ｔ４’は、予め実験的に求めておくことができる。次に、論理和回路５００で、濃度２値化回路３００の出力と、濃度変化２値化回路４００の出力との論理和をとる。以上の処理を、全ての画素に適用して、対象領域を１、背景領域を０とする２値画像を得る。このようにして得られた２値画像に、さらにフィルタ部６００で、太め処理等による対象領域の穴埋め、細め処理等による背景領域のノイズ除去の処理を行なうフイルタリングを施して、２値画像の修正を行ない、修正された２値画像が、出力部７００よりデータバス１１０に出力される（図２４）。
【０１８７】
本実施例での処理の工程を整理すると、以下のようになる。
【０１８８】
（１）濃度補正処理：Ｇ画像で、光学系の歪みに由来する濃度むらを除去する。
【０１８９】
（２）領域分割：Ｇ画像での、濃度と、濃度変化を表わす量とから、Ｇ画像を背景領域と対象領域とに分割し、背景領域を０、対象領域を１とする２値画像を作成する。
【０１９０】
（３）修正処理：対象領域の穴埋め、背景領域のノイズ除去等の２値画像の修正、整形を行なう。
【０１９１】
（４）ラベリング：２値画像の連結成分毎にラベリング処理を行ない、画像中の複数の対象に番号をつける。
【０１９２】
（５）特徴量算出：番号をつけたそれぞれの対象につき、面積、周囲長、平均濃度等の特徴量を求める。
【０１９３】
（６）識別：求めた特徴量をもとに、それぞれの対象が、どのような尿沈渣成分であるかを判断し、分類する。
【０１９４】
なお、（３）〜（６）では、太め処理、細め処理等のフィルタ処理を含め、周知の従来技術を使用できる。
【０１９５】
（実施例３）
本実施例では、実施例２の構成において、Ｇ画像の濃度に閾値を設けて領域分割する構成の代わりに、実施例１の構成（Ｇ画像、及びＲ画像の濃度に閾値を設けて領域分割する構成）を使用して、より正確に、対象粒子を抽出するものである。
【０１９６】
実際の多数のサンプルに対して適用した結果によれば、実施例１の領域分割の方法では、どちらかというと、染色細胞が、無染色細胞よりも正確に抽出でき、実施例２の領域分割の方法では、Ｇ画像のみを使用するので、染色された粒子画像のＧ画像が不鮮明である場合には、染色細胞の抽出は実施例１よりも劣ることが判明した。本実施例では、Ｇ画像、及びＲ画像の濃度に閾値を設けて領域分割するので、Ｇ画像が不鮮明である場合にも、Ｒ画像の情報が利用され、実施例１、実施例２のいずれよりも、より正確に対象粒子を抽出できることになる。
【０１９７】
本実施例での処理の工程を整理すると、以下のようになる。
【０１９８】
（１）濃度補正処理：Ｇ画像、及びＲ画像で、光学系の歪みに由来する濃度むらを除去する。
【０１９９】
（２）領域分割：Ｇ画像、及びＲ画像で複数の閾値を使用して、背景領域と対象領域とに分割し、背景領域を０、対象領域を１とする２値画像を作成する。
【０２００】
（３）領域分割：Ｇ画像での、濃度変化を表わす量から、Ｇ画像を背景領域と対象領域とに分割し、背景領域を０、対象領域を１とする２値画像を作成する。
【０２０１】
（４）領域分割：（２）、（３）で得た２値画像の論理和をとる。
【０２０２】
（５）修正処理：対象領域の穴埋め、背景領域のノイズ除去等の２値画像の修正、整形を行なう。
【０２０３】
（６）ラベリング：２値画像の連結成分毎にラベリング処理を行ない、画像中の複数の対象に番号をつける。
【０２０４】
（７）特徴量算出：番号をつけたそれぞれの対象につき、面積、周囲長、平均濃度等の特徴量を求める。
【０２０５】
（８）識別：求めた特徴量をもとに、それぞれの対象が、どのような尿沈渣成分であるかを判断し、分類する。
【０２０６】
なお、（５）〜（６）では、太め処理、細め処理等のフィルタ処理を含め、周知の従来技術を使用できる。さらに、（２）、（３）の処理結果にたいして、（５）の処理を行ない、（５）の処理を省略して実行してもよい。
【０２０７】
上記の（２）、（３）の領域分割を行なう処理は、いずれを先に行なってもよいし、平行処理してもよいことは言うまでもない。
【０２０８】
対象粒子の、赤成分の画像（Ｒ画像）、緑成分の画像（Ｇ画像）、青成分の画像（Ｂ画像）から、実施例１と同様に、対象粒子の、赤成分の画像（Ｒ画像）、緑成分の画像（Ｇ画像）を選択し、さらに実施例２と同様に、緑成分の画像（Ｇ画像）を選択して、実施例２の構成の一部に、実施例１の構成を組み込んで、実施例２において説明した、最も無染色となりやすい偏平上皮細胞を含む画像の５１例に適用した結果、８例の無染色粒子を含め、５１例の画像のすべてにおいて、良好な領域分割結果を得た。即ち、実施例２の構成の一部に、実施例１の構成を組み込み、実施例１、実施例２の単独の方法では得られなかった、良好な領域分割を行なうことができた。この結果、単独の方法では、１つの領域として理想的に分割することが困難であった無染色細胞を含め、ほとんどすべての対象粒子を、１つの連結した対象領域として分割できるようになり、沈渣成分の識別の精度を向上させることができた。
【０２０９】
本実施例では、試料の染色により、十分に染色される粒子、染色されにくい粒子（十分に染色されない粒子、ほとんど染色されない粒子）が混在する場合にも、背景領域を正確に識別して、各粒子像毎に安定して領域分割でき、より正確な２値画像が得られ、この結果、対象領域の特徴量がより正確に求められ、対象粒子の分類の誤りを少なくできる。このように、染色の程度が異なる対象粒子が混在し、染色性が悪く、ほとんど染色されない対象粒子が試料に含まれる場合にも、各粒子像毎について正確な領域分割ができ、各種の沈渣成分の識別精度が向上できる。
【０２１０】
図３０、図３１は、本実施例による、実際の試料における処理結果例を示す。
【０２１１】
図３０における試料は、図６、図１４、図１６に示す例と同じ試料であり、染色が良好な細胞（染色細胞）の例を示す。図３１における試料は、図７、図１５、図１７に示す例と同じ試料であり、染色が良好でない細胞（無染色細胞）の例を示す。
【０２１２】
図３０、図３１において、２値画像２５１、２６１は、それぞれ実施例１のＧ画像の濃度値とＲ画像の濃度値に閾値を設定する方法による領域分割の結果、２値画像２５２、２６２は、それぞれ実施例２のＧ画像の濃度値の変化を表わす量に閾値を設定する方法（ｎ＝２として、差分値を使用）による領域分割の結果を示す。２値画像２５３は、２値画像２５１と２５２との論理和をとった結果を示し、２値画像２６３は、２値画像２６１と２６２との論理和をとった結果を示す。なお、これら、２値画像２５１、２６１、２５２、２６２、２５３、２６３の領域分割の結果では、修正処理（対象領域の穴埋め、背景領域のノイズ除去等の２値画像の修正、整形処理）を行なっていない。２値画像２５４、２６４は、２値画像２５３、２６３の対して、それぞれ修正処理を行なって得た画像である。２値画像２５１に対して、修正処理を行なうと図６の画像が得られ、２値画像２６１に対して、修正処理を行なうと図７の画像が得られる。
【０２１３】
実施例３の方法では、実施例１、２の単独の方法では、試料中の粒子を理想的に抽出することは困難であっても、ほぼ見落としなく画像中の粒子を抽出することが可能となる。
【０２１４】
（実施例４）
本実施例は、実施例１から実施例３で説明した領域分割の方法により得た２値画像のいずれか２つの２値画像、もしくは３つの２値画像の論理演算（論理和）により、新たな２値画像を得て、より正確に、対象粒子を抽出するものである。
【０２１５】
図２９は、本実施例における、２つの異なる方法（実施例１、及び実施例２に基づく領域分割法）により領域分割された２つの２値画像の合成を、模式的に説明する図である。実施例１に基づく領域分割法から得た２値画像２１０と、実施例２のＧ画像の濃度値の変化を表わす量から得た２値画像２２０とから、新しい２値画像２３０が合成される。この合成は、上記２つの２値画像２１０、２２０の論理和によって得られる。２値画像２１０の粒子像２１２では、２値画像２２０で得られている粒子像２２２の周辺の一部しか抽出されておらず、逆に２値画像２２０の粒子像２２１では、２値画像２１０で得られている粒子像２１１（但し、粒子像２１１の内部の一部が、背景領域として識別されている）の内部の一部しか抽出されていない。
【０２１６】
２つの２値画像２１０、２２０の論理和によって、合成された新しい２値画像２３０では、粒子像２３１、２３２が求められ、それぞれの領域分割法の欠点を互いに補いあって、より質の向上した２値画像が得られ、より正確に目的とする対象粒子が抽出できる。以上の説明では、合成の方法として論理和をとったが、画像処理する試料の処理条件（染色条件等）、試料の種類、対象粒子の種類等により、適宜設定される。
【０２１７】
以上の各実施例において、領域分割に用いる画像として、赤画像、緑画像、青画像のどの画像を選択するかの決定は、分析対象とする試料の分光特性、背景信号のレベルやバラツキ、対象粒子が存在する領域での信号のレベルやバラツキ等を考慮して、各実施例で説明した領域分割の方法に対して、それぞれ行なえばよい。さらに、目的とする対象粒子が抽出する領域分割を行なうための論理演算として、論理和をとったが、画像処理する試料の処理条件（染色条件等）、試料の種類、対象粒子の種類等により、適宜変更されることは言うまでもない。
【０２１８】
尿試料には、通常、染色液が添加され、沈渣成分のうち細胞類は、通常、青、赤、赤紫等の色に染色される。背景は白色をしているため、単純に、背景と沈渣成分との色の違いに着目し、赤画像、緑画像、青画像、それぞれの画像信号に閾値を設け、白色に近い部分を背景領域、それ以外の領域を、目的とする粒子が存在する対象領域とする、従来の領域分割法では、染色が十分にできなかった粒子は、正確に検出することはできない。その理由は、細胞のうち、完全に死んでいない、活性度の高い細胞は、染色剤を吸収しにくい性質をもち、染色されず透明であるため、画像中の細胞の存在する領域は、背景領域とほぼ同じ色調となっているため、従来の単純な領域分割法では、正確に細胞領域を抽出できない。
【０２１９】
十分に染色されない細胞は、無色透明であり、細胞内部、及び細胞表面の不均一性により、光の屈折、散乱が生じ、細胞が存在する領域の画像の濃度値の変化は大きい。このため、単純な領域分割法では、本来、１つの細胞の存在する領域が、複数の小さな領域として分割されてしまい、それぞれの独立した領域が、異なる沈渣成分として誤って識別されてしまい、沈渣成分の識別の精度を低下させてしまう。
【０２２０】
本発明では、このような細胞像の濃度値の位置による細かな変化に着目して、画像中の各点における濃度の変化を表す量を計算し、濃度値に閾値を設けると同時に、濃度の変化を表す量に閾値を設け、領域分割を行うことにより、対象粒子が十分に染色されない場合でも、正確な領域分割を可能にするものである。
【０２２１】
（実施例５）
実施例１から実施例４において説明した方法が適用される、粒子画像の領域分割を行なう装置を、図３２を使用して以下に説明する。
【０２２２】
ＴＶカメラ１０２からの、所定の撮影領域（図３）の、Ｇ画像、Ｒ画像、及びＢ画像の画像信号は、Ａ／Ｄ変換器１０３により、例えば５１２×５１２画素に関するデジタルの画像データに変換される。これらの画像データの各画素に対して、以下に説明する各処理が実行され、粒子が存在する領域と、背景領域とが識別され、領域分割された２値画像が得られる。
【０２２３】
濃度補正回路７２は、補正用データメモリ７０に予め記憶されている、撮像光学系の歪みによる濃度むらを補正する補正用データを使用して、（数２）、又は（数３）に基づいて、濃度むらの補正を実行する。この補正結果は、画像メモリ９０に格納される。以上の構成は、実施例１から実施例３を実行する装置に共通の構成である。
【０２２４】
以下、実施例１から実施例４を実行する装置構成の例を説明する。
【０２２５】
（ａ） 実施例１を実行する装置は、図３２に示す構成のうち、ヒストグラム計算回路７４と、閾値計算回路７６と、濃度２値化回路７８と、２値画像メモリ８０とを有している。ヒストグラム計算回路７４は、濃度補正回路７２により、濃度むらが補正されたＲ画像データ、Ｇ画像データを使用して、Ｇ画像、及びＲ画像のヒストグラムを求める。閾値計算回路７６は、（数５）から（数８）に基づいて、閾値Ｔ１、Ｔ２、Ｔ３、Ｔ４をそれぞれ算出する。濃度２値化回路７８は、これら閾値Ｔ１、Ｔ２、Ｔ３、Ｔ４を使用して、Ｇ画像、及びＲ画像データから、（数９）に基づいて背景領域を検出して２値化された画像データを、２値画像メモリ８０に格納する。実施例１を実行する装置では、濃度２値化回路７８、２値画像メモリ８０は、それぞれ２値化回路９６、２値画像メモリ９８を兼ねている。（ｂ） 実施例２を実行する装置の構成は、図３２に示す構成のうち、画像メモリ９０から読み出した濃度むらが補正されたＲ画像データにおける画素間での濃度の変化の大きさを検出して２値化する濃度変化２値化回路９２と、その結果を記憶する２値画像メモリ９４と、上記（ａ）で説明した濃度むらが補正されたＲ画像、Ｇ画像データのうち、Ｇ画像データを使用する、ヒストグラム計算回路７４と、閾値計算回路７６と、濃度２値化回路７８と、２値画像メモリ８０とを有している。濃度変化２値化回路９２の具体的な構成は、実施例２においてすでに詳細に説明した、図２５、図２６、図２７、及び図２８のいずれかに示す構成と同様である。２値化回路９６は、２値画像メモリ８０と２値画像メモリ９４に格納されたデータ間での論理和（論理ＯＲ）により、領域分割を行ない、その結果を２値画像メモリ９８に格納する。
【０２２６】
なお、ヒストグラム計算回路７４と、閾値計算回路７６と、濃度２値化回路７８とからなる処理は、濃度変化２値化回路９２の処理と平行して実行されており、高速化が可能である。
【０２２７】
以上では、Ｇ画像、及びＲ画像データの濃度を使用する場合について説明したが、この構成に限定されることなく、２つの画像データは、Ｒ、Ｇ、Ｂ画像データから適宜選択される。選択された画像データを使用して、閾値計算回路７６により、複数の閾値が算出され、濃度２値化回路７８により２値化された画像データを、２値画像メモリ８０に格納する。
【０２２８】
さらに上記では、Ｇ画像データの画素間での濃度の変化の大きさを検出する濃度変化２値化回路９２と、検出された結果を格納する２値画像メモリ９４の構成について説明したが、濃度変化２値化回路９２と２値画像メモリ９４からなる構成を、さらに１、もしくは２つ並列に接続配置して、Ｇ画像データの他に、Ｂ画像データ又はＲ画像データの１画像データの、もしくはＢ画像データ、Ｒ画像データの２画像データの、画素間での濃度の変化の大きさを、平行して検出する構成とし、このように並列に接続配置された複数の濃度変化２値化回路９２からそれぞれ検出され、複数の２値画像メモリ９４に格納されている画像と２値画像メモリ８０に格納されている画像との間で、論理演算、例えば論理和を２値化回路９６で実行し、得られる２値化された画像を２値画像メモリ９８に格納する構成としてもよい。
【０２２９】
このような構成によれば、Ｒ、Ｇ、Ｂ画像データの中から複数の画像を使用して、それぞれの画像において、画素間での濃度の変化の大きさを検出して、２値化された画像データを得るので、Ｒ、Ｇ、Ｂ画像データの中の単独の画像を使用し、画素間での濃度の変化の大きさを検出して、２値化された画像データを得る場合に比較して、より正確に粒子が存在する粒子領域を識別できる。
【０２３０】
（ｃ） 実施例３を実行する装置は、図３２に示す構成を有している。即ち、上記（ａ）、及び（ｂ）の構成が結合された構成を有している。即ち、画像メモリ９０から読み出した濃度むらが補正されたＧ画像データにおける、画素間での濃度の変化の大きさを検出して２値化された画像データを得る濃度変化２値化回路９２と、２値化された画像データを格納する２値画像メモリ９４と、濃度むらが補正されたＲ画像データ、Ｇ画像データを使用して、Ｇ画像、及びＲ画像のヒストグラムを求めるヒストグラム計算回路７４と、Ｇ画像、及びＲ画像データから、閾値Ｔ１、Ｔ２、Ｔ３、Ｔ４を使用して、（数９）に基づいて背景領域を検出して２値化された画像データを得る濃度２値化回路７８と、２値化された画像データを格納する２値画像メモリ８０と、２値画像メモリ８０と２値画像メモリ９４に格納されたデータ間での論理和（論理ＯＲ）演算を行なう２値化回路９６と、論理和（論理ＯＲ）演算の結果であり、領域分割された結果を格納する２値画像メモリ９８と、から構成される。なお、濃度変化２値化回路９２の具体的な構成は、実施例２においてすでに詳細に説明した、図２５、図２６、図２７、及び図２８のいずれかに示す構成と同様である。
【０２３１】
なお、ヒストグラム計算回路７４と、閾値計算回路７６と、濃度２値化回路７８とからなる処理は、濃度変化２値化回路９２の処理と平行して実行されており、高速化が可能である。
【０２３２】
以上では、Ｇ画像、及びＲ画像データの濃度を使用する場合について説明したが、この構成に限定されることなく、２つの画像データは、Ｒ、Ｇ、Ｂ画像データから適宜選択される。選択された画像データを使用して、閾値計算回路７６により、複数の閾値が算出され、濃度２値化回路７８により２値化された画像データを、２値画像メモリ８０に格納する。また、ヒストグラム計算回路７４と、閾値計算回路７６と、濃度２値化回路７８、２値画像メモリ８０とからなる構成を複数個並列接続して、Ｇ画像、及びＲ画像データの濃度にさらに、例えばＧ画像、及びＢ画像データの濃度を使用して、複数の２値画像メモリ８０に２値化された画像を得て、２値画像メモリ８０に格納されている画像と、複数の２値画像メモリ８０に格納されている画像との間で、論理演算、例えば論理和を２値化回路９６で実行し、得られる２値化された画像を２値画像メモリ９８に格納する構成としてもよい。
【０２３３】
さらに、上記では、Ｇ画像データの画素間での濃度の変化の大きさを検出する濃度変化２値化回路９２と、検出された結果を格納する２値画像メモリ９４の構成について説明したが、濃度変化２値化回路９２と２値画像メモリ９４からなる構成を、さらに１、もしくは２つ並列に接続配置して、Ｇ画像データの他に、Ｂ画像データ又はＲ画像データの１画像データの、もしくはＢ画像データ、Ｒ画像データの２画像データの、画素間での濃度の変化の大きさを、平行して検出する構成とし、このように並列に接続配置された複数の濃度変化２値化回路９２からそれぞれ検出され、複数の２値画像メモリ８０に格納されている画像と、１又は複数の２値画像メモリ９４に格納されている画像との間で、論理演算、例えば論理和を２値化回路９６で実行し、得られる２値化された画像を２値画像メモリ９８に格納する構成としてもよい。
【０２３４】
このような構成によれば、Ｒ、Ｇ、Ｂ画像データの中から複数の画像を使用して、それぞれの画像において、画素間での濃度の変化の大きさを検出して、２値化された画像データを得るので、Ｒ、Ｇ、Ｂ画像データの中の単独の画像を使用し、画素間での濃度の変化の大きさを検出して、２値化された画像データを得る場合に比較してより正確に粒子が存在する粒子領域を識別できる。
【０２３５】
（ｄ） 実施例４を実行する装置の構成は、上記で説明した（ａ）、（ｂ）の装置構成を備えており、（ａ）、（ｂ）で説明した装置構成部分から得られた結果を、２値化回路９６によって、論理演算（論理和）をとり、領域分割された２値画像を得て、２値画像メモリ９８に格納する。
【０２３６】
以上説明した（ａ）から（ｄ）の装置構成の２値画像メモリ９８に得られた、領域分割された２値画像を使用して、粒子が存在する対象領域に関して特徴量が、演算処理手段（図３２には図示せず）によって演算検出され、尿沈渣粒子の識別が実行される。
【０２３７】
【発明の効果】
本発明によれば、特に、変化に富んだ性質をもつ対象粒子の検出が要求される尿沈渣検査装置で得る画像の領域分割において、試料の染色により、対象粒子が十分に染色されず、対象粒子の濃度が背景濃度とほとんど差がなく、このような対象粒子を高い頻度で多数含む試料の場合にも、対象粒子の抽出ができる。即ち、試料の染色により、十分に染色される粒子、染色されにくい粒子（十分に染色されない粒子、ほとんど染色されない粒子）が混在する場合にも、各粒子像毎に安定して分割でき、より正確な２値画像が得られ、この結果、対象領域の特徴量がより正確に求められ、対象粒子の分類の誤り防止が可能となり、各種の沈渣成分の識別精度が向上できる。
【図面の簡単な説明】
【図１】本発明の第２の実施例の粒子画像の像領域分割方法の原理を説明する図。
【図２】本発明の第２の実施例の粒子画像の像領域分割方法の原理を説明する図。
【図３】本発明の領域分割方法が適用される尿沈渣検査装置の構成例を示す図。
【図４】本発明の色空間での背景部の抽出を行なう領域分割を模式的に示す図。
【図５】本発明での、閾値の決定に使用するパラメータを求めるための濃度ヒストグラムを説明する図。
【図６】本発明の実施例１での、対象粒子の染色が良好な染色細胞の画像、対象粒子の染色がほとんどされていない細胞の画像、の領域分割結果の例を示す図。
【図７】本発明の実施例１での、対象粒子の染色が良好な染色細胞の画像、対象粒子の染色がほとんどされていない細胞の画像、の領域分割結果の例を示す図。
【図８】本発明の実施例２での、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の濃度値の例を示す図。
【図９】本発明の実施例２での、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の濃度値の例を示す図。
【図１０】本発明の実施例２での、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の分散値の例を示す図。
【図１１】本発明の実施例２での、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の分散値の例を示す図。
【図１２】本発明の実施例２での、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の濃度差分値の例を示す図。
【図１３】本発明の実施例２での、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の濃度差分値の例を示す図。
【図１４】本発明の実施例２での、濃度の局所的分散を用いた２値化による方法を用いて行なった、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の領域分割の結果例を示す図。
【図１５】本発明の実施例２での、濃度の局所的分散を用いた２値化による方法を用いて行なった、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の領域分割の結果例を示す図。
【図１６】本発明の実施例２での、濃度差分値を用いた２値化による方法を用いた、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の領域分割の結果例を示す図。
【図１７】本発明の実施例２での、濃度差分値を用いた２値化による方法を用いた、染色が良好な細胞を含む画像、染色がほとんどされていない細胞を含む画像、の領域分割の結果例を示す図。
【図１８】本発明の実施例２での、マスクサイズの変化による領域分割の結果例を示す図。
【図１９】本発明の実施例２での、１走査線における背景領域、及び対象領域における、離散フーリエ変換した結果例を示す図。
【図２０】本発明の実施例２での、フイルタの周波数振幅特性（｜Ｈ（Ω）｜）の例を示す図。
【図２１】本発明の実施例２での、染色された赤血球の画像を用いて、分割される領域の形状のマスクサイズによる影響を調べた結果例を示す図。
【図２２】本発明の実施例２での、粒子画像の領域分割の手順を説明する処理のフローチャート。
【図２３】本発明の実施例２を実施するための装置の構成例を示す図。
【図２４】図２３における領域分割部の構成例を示す図。
【図２５】本発明の実施例２での、濃度変化計算部の具体的な回路構成例を示す図。
【図２６】本発明の実施例２での、濃度変化計算部の具体的な回路構成例を示す図。
【図２７】本発明の実施例２での、濃度変化計算部の具体的な回路構成例を示す図。
【図２８】本発明の実施例２での、濃度変化計算部の具体的な回路構成例を示す図。
【図２９】本発明の実施例４での、２つの異なる方法により領域分割された２つの２値画像の合成を、模的に説明する図。
【図３０】本発明の実施例３での、実際の試料における領域分割の結果例を示す図。
【図３１】本発明の実施例３での、実際の試料における領域分割の結果例を示す図。
【図３２】本発明の粒子画像の像領域分割方法が適用される装置の構成例を示す図。
【符号の説明】
２２…尿試料、２４…染色液、２６…フローセル、２８…パーソナルコンピュータ、３０…フラッシュランプ（Ｘｅランプ）、３２…半導体レーザ、３４…粒子検出器、３６…ＴＶカメラ、４０…Ｒ画像の濃度ヒストグラム、４０’…Ｇ画像の濃度ヒストグラム、５０、５１、５２…細胞、６０、６１、６２…ノイズ成分、７０…補正用データメモリ、７２…濃度補正回路、７４…ヒストグラム計算回路、７６…閾値計算回路、７８…濃度２値化回路、８０…２値画像メモリ、９０…画像メモリ、９２…濃度変化２値化回路、９４…２値画像メモリ、９６…２値化回路、９８…２値画像メモリ、１０２…ＴＶカメラ、１０３…Ａ／Ｄ変換器、１０１…パルスランプ、１０２…テレビカメラ、１０３…Ａ／Ｄ変換器、１０４…画像メモリ、１０５…閾値計算部、１０６…領域分割部、１０７…画像メモリ、１０８…計算機、１０９…ＣＰＵバス、１１０…データバス、２００…入力部、２１０、２２０、２３０…２値画像、２１１、２１２、２２１、２２２、２３１、２３２…粒子像、３００…濃度２値化回路、３１０…閾値メモリ、３２０…比較器、４００…濃度変化２値化回路、４１０…メモリ、４１１〜４１４…、４２０…濃度変化計算部、４２１、４２２…加算器、４２３…減算器、４３０…比較部、４３１…閾値メモリ、４３２…比較器、５００…論理和回路、６００…フィルタ部、７００…出力部、８００…分散計算回路、８１１〜８１９…メモリ、８２１…２乗演算部、８２２、８２４…加算器、８２３、８２５…除算器、８２６…２乗演算器、８２７…減算器、９００…荷重和計算回路、９３１〜９３９…メモリ、９４１〜９４９…重み値メモリ、９５１〜９５９…乗算器、９６１…加算器、１０００…画像メモリ。[0001]
[Industrial application fields]
The present invention relates to an image area dividing method using color information and density information, and more particularly to an area dividing method and apparatus suitable for dividing a particle image in blood or urine.
[0002]
[Prior art]
Conventionally, in order to morphologically examine particles in urine, a visual method has been performed by centrifuging a urine sample, staining a sediment, preparing a specimen on a slide glass, and observing under a microscope. . At that time, the kind of the sediment and its concentration in the original urine sample were determined with the degree of concentration of the sample by centrifugation and the amount of the sample to be observed constant.
[0003]
Some sediments range from particles of a few micrometers, such as blood cells and bacteria, to particles of a few hundred micrometers, such as a cylinder. These particles with different sizes were observed by switching.
[0004]
In the conventional visual inspection by microscopic observation, there is a problem that (1) a method using a slide glass has a low throughput, and (2) it is difficult to handle a urine sample that is quickly rotted.
[0005]
In recent years, the blood image test is shifting from a system using a slide glass specimen to a flow system in which a blood sample is directly tested as a liquid sample. The flow method is expected to realize high-speed sample processing.
[0006]
Attempts to take images of particles in a continuously flowing sample specimen and analyze and classify the particles from the individual particle images are described in JP-T-57-500995 and JP-A-63-94156. ing.
[0007]
In Japanese Patent Publication No. 57-500995, a sample specimen is passed through a specially shaped flow path, particles in the specimen are allowed to flow through a wide imaging area, and a still image is taken using a flash lamp. A method for particle analysis using images is described. When a magnified image of sample particles is projected onto a CCD camera using a microscope, a flash lamp, which is a pulse light source, periodically emits light in synchronization with the operation of the CCD camera. The light emission time of the pulse light source is short, a still image can be obtained even if particles flow continuously, and the CCD camera can capture 30 still images per second.
[0008]
Japanese Patent Application Laid-Open No. 63-94156 discloses a separate particle detection optical system for detecting the passage of particles upstream of the particle image capturing region of the sample flow, in addition to the still image capturing system. In this method, the passage of particles is detected in advance by the particle detection unit, and when the particles reach the particle imaging region, the flash lamp is turned on at an appropriate timing. In this method, the pulsed light source is not periodically emitted, the passage of particles is detected, and the flash lamp is turned on only at that time, and a still image can be taken. As a result, particle images are efficiently collected, and in the case of a sample sample having a low concentration, a still image is not photographed in a portion of the sample flow where particles are not present, and therefore unnecessary image processing is not required.
[0009]
Japanese Patent Application Laid-Open No. 5-296915 describes a method in which a particle detection optical system for detecting particles is incorporated in a particle image pickup system. That is, a method of irradiating a sample flow with a laser beam for particle detection for detecting particles through a microscope condenser lens of a microscope image pickup system is described. In this method, it is not necessary to prepare a separate particle detection optical system for detecting particles, and the particle detection position can be arranged as close to the particle image capturing region as possible.
[0010]
As a prior art of the region dividing method of the image thus obtained, for example, Japanese Unexamined Patent Application Publication No. 1-119765 discloses a region dividing method for classifying blood cell images. In this technique, a region of an image is divided in a color space using a threshold value obtained from an image density histogram.
[0011]
[Problems to be solved by the invention]
  However, when the density distribution of the particle image to be analyzed is a density distribution that is widely distributed from a low density value to a high value in an image to be subjected to area division, the conventional technique cannot always perform accurate area division. There was a problem. In addition, the concentration distribution characteristics of the analysis targetParticulateBecause it depends on the color tone, DoubleNumber of particle imagesRegion divisionIf the same image processing is performed, there is a problem that the area is not necessarily accurate.
[0012]
In urine sediments, samples are usually stained to prevent misidentification of sediment components and facilitate cell determination. There are several known methods for taking an image of a stained biological sample and extracting target particles in a color space. However, in the urine sediment examination apparatus, the target particles to be divided into regions have a variety of properties as follows.
[0013]
(1) The range of the size distribution of the target particles is as wide as several μm to several hundred μm.
[0014]
(2) A plurality of particles having different color tones are present in the same image.
[0015]
(3) The concentration of target particles may change from a small value to a large value.
[0016]
(4) Even with the same type of cells, there are cells that are darkly stained and cells that are lightly stained.
[0017]
(5) The concentration of the target particles may be the same as or lower than the concentration of the background portion.
[0018]
(6) Particles that are sufficiently dyed and particles that are difficult to be dyed (particles that are not dyed sufficiently, particles that are hardly dyed) coexist by staining the sample.
[0019]
As described above, since the target particles in the urine sediment examination apparatus have a variety of properties, there has been a problem that the conventional region segmentation method cannot be simply applied.
[0020]
In particular, when the target particles are not sufficiently stained, there is almost no difference between the cytoplasm concentration and the background concentration, and there is a problem in that it cannot be extracted as an accurate binary image in which all the regions where the target particles exist have the same value. It was.
[0021]
Furthermore, cells with high activity have poor staining properties, and there are cells that are hardly stained, and in images containing such cells, the target area where the target particles are present and the background area have almost the same color tone, and the target area Since there are many portions having the same density as the background region, there is a problem that such cells cannot be extracted accurately.
[0022]
As described above, when the target particles are not sufficiently (satisfactorily) dyed and an accurate binary image cannot be obtained, the feature amount of the target region cannot be obtained accurately, and the target particles are misclassified. There was a problem.
[0023]
Since target particles such as cells that are poorly stained and hardly stained are generated at a high frequency, the conventional technique cannot always accurately divide all the particles contained in the sample, and the accuracy of sediment component identification is not high. There was a problem that hindered improvement.
[0024]
In order to solve the above problems, in a region segmentation method in which particles scattered in a fluid are captured and recorded as a still image, and target target particles are extracted from the still image, the first object of the present invention An object of the present invention is to provide a region dividing method that can stably extract and accurately extract a particle image having a difference in density and a particle image having a color tone difference from a still image.
[0025]
The second object of the present invention is to generate a particle image including a large number of regions having a density similar to that of the background region in the target region where the target particle is present and the color difference from the background region is small. Even in this case, an object of the present invention is to provide a region dividing method capable of dividing a particle image stably and accurately.
[0026]
The third object of the present invention is to provide a background image in a target area where there is a small difference in color tone from a static image, a particle image having a density difference, a particle image having a color tone difference, and a background area. An object of the present invention is to provide a region dividing method that can stably and accurately divide each particle image even when particle images including a large number of regions having the same or lower density than the regions are mixed.
[0027]
[Means for Solving the Problems]
  The above object is achieved by the following configuration. That is, a red component image (R image), a green component image (G image), and a blue component image (B image) of a stained target particle obtained by an image input optical system that inputs a particle image. Particles that are captured and stored as still images, and that use one or more of the R image, the G image, and the B image to identify a background region and a target region where the target particles exist An image region segmentation method, wherein the background of the first group of one or more images selected from the R image, the G image, and the B image with respect to the density value of each image A first step of obtaining a binarized first binary image from the first group image by setting a first group threshold value for dividing the region and a region having a large density difference; and It consists of one or more selected from the R image, the G image, and the B image. An amount representing the magnitude of the change in the density value in the vicinity region of each pixel constituting each image of the second group of images is calculated for each pixel, and the density obtained for each pixel is calculated. A threshold value of the second group for dividing the region having a large change in the density value is set as an amount representing the magnitude of the change in the value in comparison with the change in the density value in the background region. A second step of obtaining a binarized second binary image from the image, a logical operation of the first binary image and the second binary image, and the background region; And a third step of obtaining a binarized third binary image representing the target region.
Hereinafter, each component will be described. First, the configuration for identifying the background region and the target region based on the density value, which is a part of the configuration of the present invention, will be specifically described. <1> That is, the aboveThe first purpose is that the dyed sample is separated by wavelength by the image input optical system for inputting an image from the stream in which the dyed sample flows together with the fluid, and the target dyed in the stream is stained. The red component image (R image), the green component image (G image), and the blue component image (B image) of each target particle are captured and stored as still images, and the R image, the G image, and the B image are stored. A particle image region segmentation method that uses a plurality of images in an image to identify a background region and a target region where target particles exist, and includes at least one of an R image, a G image, and a B image. A step of selecting two images and correcting density unevenness due to distortion of the image input optical system included in each of the two images, two images, and a plurality of threshold values (T1, T2, T3, T4) ) To identify the background area and from the two images Obtaining a binarized binary image representing a background region and a region other than the background region and containing the target particle, and extracting the target region where the target particle exists Is achieved by a region segmentation method characterized by With this configuration, target particles that are target sediment components can be extracted from the above-described image converted into a binary image.
[0028]
By correcting the density unevenness due to the distortion of the image input optical system, the deviation of the density fluctuation in the at least two images is reduced, and the background region can be stably identified and accurately extracted. It is preferable to select a G image and an R image as the two images. In the image obtained by the urine sediment examination apparatus, it is known that the background region usually exists stably, and therefore the target region where the target particles are present can be identified on the basis of the background region. That is, since the background region can be stably extracted, it can be considered that the signal components representing the target particles to be detected separately from the background region are included in the region other than the background region.
[0029]
  The process of converting to the above binary image will be described below by taking a specific case where a G image and an R image are selected as the two images as an example. A density histogram of each of the G image and the R image is obtained, and a density value having the maximum peak value of the histogram is obtained. Next, threshold values T1 and T2 are obtained from the density histogram of the R image, and threshold values T3 and T4 are obtained from the density histogram of the G image.TheObtain the background area.
  Below, the outline | summary of the procedure which calculates | requires each threshold value and extracts a background area | region is demonstrated.
[0030]
(Procedure 1) A density histogram of each image (G image, R image) is created.
[0031]
(Procedure 2) A density value (maximum density value) having the maximum frequency value of each density histogram is obtained.
[0032]
(Procedure 3) The density giving the position of the half width of the peak of each density histogram is obtained.
[0033]
(Procedure 4) Threshold values T1, T2, T3, and T4 are calculated from the maximum density value, the density that gives the position of the half-width of the peak, and the predetermined parameters obtained in (Procedure 1) to (Procedure 3). To do.
[0034]
(Procedure 5) A binary image is created in the image space based on the threshold values T1, T2, T3, and T4 calculated in (Procedure 1) to (Procedure 4). The background area is obtained by (Equation 1).
[0035]
{T1 ≦ R (i, j) <T2} ∩ {T3 ≦ G (i, j) <T4} (Equation 1)
Here, ∩ represents the logical product, (i, j) represents the pixel position, and R (i, j) and G (i, j) were each corrected for density unevenness due to distortion of the image input optical system. R image and G image are represented.
[0036]
<2>Next, a configuration for identifying the background region and the target region based on the magnitude of the density value change, which is a part of the configuration requirement of the present invention, will be described. That is, the aboveA second configuration for realizing the second object will be described below. The second purpose is that the dyed sample is separated by wavelength by the image input optical system that inputs an image from the flow in which the dyed sample flows along with the fluid. The red component image (R image), the green component image (G image), and the blue component image (B image) of each target particle are captured and stored as still images, and the R image, the G image, and the B image are stored. This is a particle image region segmentation method that uses one image in an image to identify a background region and a target region in which target particles exist, and is a predetermined method from among R, G, and B images. A step of correcting the density unevenness caused by distortion of the image input optical system included in the predetermined image, and an area having a large density difference between the background area and the density value of the predetermined image. A first group of thresholds for dividing (first and second thresholds) To obtain a binarized first binary image, and the amount representing the magnitude of the change in density value in the vicinity region of each point constituting the predetermined image Compared with the change in the density value in the background area, the first value for dividing the area having the large change in the density value into an amount representing the magnitude of the change in the density value obtained at each point. Setting two groups of threshold values (third threshold value, or fourth and fifth threshold values) to obtain a binarized second binary image; a first binary image; A binarized third binary image representing a background region and a target region that is a region other than the background region and includes target particles is obtained by a logical operation with the binary image of 2, for example, a logical sum. A region dividing method comprising: detecting a background region from a third binary image, and detecting a target region. More is achieved.
[0037]
The amount representing the magnitude of the change in density value is, for example, (1) the difference between the sums of density values in two adjacent small regions in the image, or the absolute value thereof, (2) ▼ Dispersion value or standard deviation of density value distribution in small area in image, (3) Load sum of density value of pixels included in small area in image, or absolute value thereof, or (4) ▲ Any one of the values (1) to (3) (amount representing the magnitude of the change in the density value) is newly regarded as a density value, and the process of applying the methods (1) to (3) is performed once again. The value obtained by performing the operation once or a plurality of times is used. It is preferable to select a G image as the predetermined image.
[0038]
<3> A third configuration for realizing the third object of the present invention will be described below. The third purpose is that the dyed sample is separated by wavelength by the image input optical system that inputs an image from the flow in which the dyed sample flows along with the fluid. The red component image (R image), the green component image (G image), and the blue component image (B image) of each target particle are captured and stored as still images, and the R image, the G image, and the B image are stored. A particle image region segmentation method that uses a plurality of images in an image to identify a background region and a target region where target particles exist, and includes at least one of an R image, a G image, and a B image. Select two images, correct the density unevenness due to distortion of the image input optical system included in the two images, select one of the two images, and select the image The darkness in the vicinity of each point constituting a given image The amount representing the magnitude of the change in the value is calculated at each of the above points, and the amount representing the magnitude of the change in the density value obtained at each point is compared with the change in the density value in the background region. A process for obtaining a binarized first binary image by setting a first threshold value (first threshold value, or second and third threshold values) for dividing an area having a large value change.AboutA plurality of threshold values (T1, T2, T3, T4) are provided in the density values of the two images and the density values of the two images, the background region is identified in the predetermined image, and the predetermined image is And a second binarized image representing the target area where the target particles existAboutStep C, the first binary image, and the second binary imageWithA process to obtain a binarized third binary image by logical operation (logical sum)AboutAnd the step D. The method is achieved by a region dividing method characterized by detecting a target region in which target particles exist from the third binary image.
[0039]
In step A, it is preferable to select a G image and an R image as the two images, and in step B, select a G image as the predetermined image. It goes without saying that either of the execution of Step B and Step C may be executed first. Of course, it is more preferable that the processing time is shortened by executing Step B and Step C in parallel at the same time.
[0040]
<4> A fourth configuration for realizing the third object of the present invention will be described below. A new region division result is obtained from any two binary images or three binary image logical operations (logical sums) of the binary images obtained from the first to third configurations. In the first configuration, as the predetermined image, two or three different images are sequentially selected and executed from the G image, the R image, and the B image, and a logical operation of a plurality of third binary images obtained is obtained. For example, a configuration may be adopted in which a target region where target particles exist is detected from a logical sum.
[0041]
[Action]
  In the first configuration of the present invention, according to the above procedure, the distribution of the concentration of the target particle image in the image to be divided (image of the stained sample) is from a low value to a high value. Even when the concentration distribution is widely distributed, accurate region division can be performed. Further, the region can be divided even when the density distribution characteristic of the analysis object varies depending on the color tone. In other words, when extracting target particles of interest from a stained sample image, even a particle image with a difference in density or a particle image with a difference in color tone can be stably and accurately affected by the difference in density and color difference. Can be divided into regions, and as a result,Of multiple particlesAreaDiscountEven if the same image processing is performed, it can be recognized as an accurate region.
[0042]
In the second configuration of the present invention, when an amount that takes only a positive value is used as an amount (a value obtained by the methods (1) to (4) above) representing the magnitude of the change in density value, If the horizontal axis represents the density value at each point of the image and the vertical axis represents the amount of change in the density value, the background area is BG, and the object having a higher density value than the background area, as shown in FIG. Is distributed as shown by A1, A2 is a target having a density value lower than that of the background area, and B is a target whose density difference is small compared to the background area and whose density value is larger than that of the background area. . Therefore, the threshold value T1 'and T2' can be set for the density value, and the threshold value T3 'can be set for the magnitude of the density value change, so that the image can be divided into the background area and the target area.
[0043]
Further, when an amount that can take a negative value is used as the amount representing the magnitude of the change in the density value, as shown in FIG. 2, the background area is BG, and the object having a density value higher than the background area is A1. An object having a density value lower than that of the background area is distributed as indicated by A2, and objects whose density difference from the background area is low and whose density value changes compared to the background area are indicated by B1 and B2, respectively. . In this case, the background area and the target area can be divided by setting the threshold values T1 ′ and T2 ′ for the density value and the threshold values T3 ″ and T4 ′ for the magnitude of the change in the density value. When this method is used, the target region can be extracted more accurately by using the threshold value of the magnitude of the density value change.
[0044]
In addition, according to the second configuration of the present invention, in the urine sediment inspection apparatus that requires region division of target particles having various properties as described above, the target particles are sufficiently obtained by staining the sample. The target particles can be extracted even in the case of a sample that is not stained and the concentration of the target particles is almost the same as the background concentration and includes a large number of such target particles. The sample contains cells that are hardly stained, the color difference between the target area where the target particles are present and the background area is small, and within the target area where the target particles are present, the density of the background area is the same as the background area. Even when there are many parts (regions) with a density lower than the density, such target particles can be stabilized for each particle image by simple calculation without reducing the processing speed of target particle extraction. Can be extracted accurately and divided into regions.
[0045]
In the third and fourth configurations of the present invention, each particle image is obtained even when particles that are sufficiently dyed and particles that are difficult to dye (particles that are not dyed sufficiently, particles that are hardly dyed) are mixed due to the dyeing of the sample. Each image can be divided stably and a more accurate binary image can be obtained. As a result, the feature amount of the target region can be obtained more accurately, and the target particle classification error can be prevented. In this way, even when target particles with different degrees of staining are mixed, target particles that are poorly dyeable and hardly stained are included in the sample, accurate segmentation can be performed for each particle image, and various sediment components It has the feature that the identification accuracy of can be improved.
[0046]
【Example】
First, urine sediment examination to which the method of the present invention can be preferably applied will be described.
[0047]
In urine sediment, samples are often stained to prevent misidentification of sediment components and facilitate cell determination. As typical staining methods, there are Sternheimer-Malbin (SM) staining and Sternheimer (S) staining modification method. SM staining is widely used mainly for staining of leukocytes. On the other hand, a modified S staining method (hereinafter abbreviated as S staining) is a staining method using Alcian blue or Astral blue as a basic dye and pyronin B as a basic dye. This staining method has an advantage that the contrast between red and blue is increased because the cell nucleus and the columnar substrate are clearly blue-stained and the cytoplasm and RNA components are red-stained by pyronin B. In addition, various dyeing methods are known. In the urine sediment test, the stained sample is used to classify and count blood cells and columnar epithelial cells, identify the type of crystals, identify the presence of bacteria including yeast-like fungi, and determine the type of bacteria. Etc. are executed.
[0048]
In the following embodiments, as an example, image segmentation in a urine sediment examination apparatus that automatically analyzes solid matter (urine sediment) in urine stained by a modified Sternheimer staining method will be described.
[0049]
In each of the following examples, the stained sample is flowed to the flow cell forming a flat sheath flow without being centrifuged from the urine sample, and the flow cell is irradiated with pulsed light in synchronization with the sediment particles flowing through the flow cell. Then, a still image is obtained, captured by a TV camera, and urinary sediments are identified and classified by image processing. In the image processing described in each of the following examples, the identification of a plurality of sediments (inspection target particles) present in the recorded image is matched with the imaging cycle (1/30 second) of the TV camera at the maximum. Thus, the image processing can be performed at high speed.
[0050]
FIG. 3 shows a configuration example of the urine sediment examination apparatus. The urine sample 22 is mixed with the staining liquid 24 and stained, and after being shaped into a flat shape having a predetermined thickness and width in the flow cell 26, the urine sample 22 is wrapped in a sheath liquid and a space having the predetermined dimension is formed. , Flowing at the same speed. In this apparatus, the light from the semiconductor laser 32 is irradiated to the particle detection region existing in front of the imaging region, and particles passing through the particle detection region are detected by the particle detector 34, and the detection signal from the particle detector 34 is detected. After that, after a certain time delay (time until the particles enter the imaging region from the particle detection region), the flash lamp (Xe lamp) 30 emits light for a predetermined short time, and the imaging optical system (objective lens) , The relay lens), the shooting area is imaged by the TV camera 36. An output signal of the camera 36 is an A / D conversion for each wavelength of a red image, a green image, and a blue image (hereinafter abbreviated as R image, G image, and B image) by an image processing board loaded in the personal computer 28. Each image is quantized to 512 pixels × 512 pixels × 8 bits (a total of 786,432 bytes of data) and stored in the image memory 38. At this time, the output of the A / D converter (not shown) is such that the black level of the camera is 0 and the white level is 255. However, in the following processing, for the convenience of processing, the black level is set to 255 and the white level is set to 0 in order to express the target by the light absorption rate.
[0051]
In the image input system, since a lens moving mechanism for magnification enlargement is disposed between the objective lens and the camera, density unevenness due to distortion of the imaging optical system occurs. In order to correct and remove this uneven density, a density correction process is performed before the area division process. This correction is performed independently for each of the R image, the G image, and the B image. As an image to be subjected to the density correction process, an image that satisfies the most preferable condition is selected in advance according to the staining method of the sample.
[0052]
Originally, when there is no target, or when shooting an object with uniform optical characteristics with an infinitesimal light absorption rate (white image), and an image when shooting an object with an infinite light absorption rate In (black image), ideally, the value of each pixel of the image is uniform, but in reality, density unevenness of the background occurs due to distortion of the optical system. Therefore, the distortion was corrected by correcting and converting the input image f (i, j) using the actually measured white image W (i, j) and black image B (i, j). An image f ′ (i, j) is obtained. The image f is any one of an R image, a G image, and a B image, and (i, j) represents a pixel position of the image. Set the black level reference value to β₀, White level reference value α₀And the relative value of the input image f (i, j) is [β₀, Α₀], The image f ′ (i, j) whose distortion is corrected can be expressed by the following (Equation 2).
[0053]

However, i = 0, 1, 2,..., 510, 511, j = 0, 1, 2,.
[0054]
Approximately, if the variance of the black image B (i, j) is 1 level or less as a result of A / D conversion, the black image B (i, j) is averaged B_avIf the pixel in the white image W ′ (i ′, j ′) in the small area is referred to without referring to each pixel in the white image W (i, j), the following (number According to 3), the same correction as described above can be performed.
[0055]

Here, if the size of the small area of the white image to be referred to is M × N, (i ′, j ′) indicates the coordinates to be referred to,
i '= [i / M]
j ′ = [j / N] (Expression 4)
And [] is a Gaussian symbol. For example, if M = N = 8, the size of the white image W ′ may be 64 × 64 pixels.
[0056]
Black level reference value β₀, White level reference value α₀Is β for both R, G, and B images.₀= 230 and α = about 40.
[0057]
Further, the white image is subjected to smoothing processing within the image, or a plurality of white images are measured, and the average processing or smoothing processing of the image is performed to reduce the influence of noise. With this correction, the density of the photographed color image becomes almost uniform over the entire image. It goes without saying that the image correction described above can be omitted when the distortion of the optical system can be substantially ignored.
[0058]
The above description is common processing executed prior to area division in each of the following embodiments.
[0059]
Example 1
  Example 1 is an example showing a part of the configuration of the present invention, and a configuration for identifying a background region and a target region based on a density value will be specifically described.In this embodiment, in order to efficiently divide a plurality of particles having different sizes (several μm to several hundred μm) present in an image captured by the flow method, from an image input by wavelength with a TV camera, After correcting the density unevenness of the optical system, the background region and the target region where the target particle exists are divided in the color space of the red image and the green image, and a binary value representing the background region and the target region where the target particle exists Get an image.
[0060]
In the present embodiment, the sediment component region in the image is divided from the image of the sample stained with S staining having a high blue and red contrast, which is suitable for automatic identification of the inspection target particles. The area division is performed based on the background area of the image. The correction process described above is performed in advance to correct the density unevenness of the optical system, to reduce the deviation of the density fluctuation, and to stably extract the background region.
[0061]
The details of the region dividing method will be described below.
[0062]
In the region division method described below, even when a target image of a target particle of interest is extracted from an image of a sample stained with S staining, a particle image having a density difference or a particle image having a color difference is stably divided. There are features that can be done.
[0063]
There are several known methods for taking an image of a stained biological sample and extracting a target substance (particles) in a color space. However, in the urine sediment examination apparatus, the object to be subjected to area division has a variety of properties as described above, and the conventional area division method cannot be simply applied as it is.
[0064]
According to the research of the inventor of the present invention, in the image obtained by the urine sediment examination apparatus, it has been found that the background region usually exists stably, so in order to detect the target particles of interest, In the present invention, a stable background area is first identified and a method based on this is adopted. That is, if the background region can be extracted stably, it can be considered that all signal components representing target particles to be detected separately from the background region are included in the region other than the background region.
[0065]
From the result of measuring typical sediment particles with a microspectrophotometer, the absorption peak of the staining agent is in the vicinity of 550 nm, and compared with the sensitivity in the wavelength region 400 to 500 nm (blue component), the wavelength region 500 to 700 nm ( It was found that the sensitivity of the green component and the red component was high. From this result, in the case of an image of a sample stained with S staining, it is clear that the region division is preferably performed in the G image and R image spaces. FIG. 4 schematically shows area division for extracting the background portion in the color space. The x-axis is the density of the G image, and the y-axis is the density of the R image. Here, the threshold values T1 and T2 are obtained from the density histogram 40 of the R image, and the threshold values T3 and T4 are obtained from the density histogram 40 ′ of the G image (a method similar to the region division method shown in FIG. 4 is used). However, in Japanese Patent Laid-Open No. 6-314338, the uneven density due to the distortion of the imaging optical system described above is not corrected).
[0066]
  Specifically, the density histogram of each image is obtained, and the density value having the maximum peak value of the histogram, the half valuevalueThe density value giving the width is obtained. Next, based on these values, a background area necessary for extracting the target area is identified and extracted.
[0067]
Hereinafter, a procedure for calculating each threshold value, identifying a background region, and obtaining a feature amount of the identified target particle pattern will be described.
[0068]
(Procedure 1): A density histogram of each image (G image, R image) is created. The pixel position of each image is represented by (i, j), the G image is represented by G (i, j), and the R image is represented by R (i, j).
[0069]
(Procedure 2): In each density histogram, the density value P having the maximum frequency value Pmax (*) as shown in FIG._dFind (*). Note that * represents one of R and G.
[0070]
(Procedure 3): In each density histogram, the density giving the half width of the peak, P_hL(*), P_hhFind (*). * Represents either R or G.
[0071]
(Procedure 4): Density value P obtained in (Procedure 1) to (Procedure 3)_d(*), P_hL(*), P_hhUsing (*), threshold values T1, T2, T3, and T4 are respectively calculated by the following (Equation 5) to (Equation 8). From the R image,
T1 = P_d(R)-{P_d(R) -P_hL(R)} · σ (Expression 5)
T2 = P_d(R)-{P_d(R) -P_hh(R)} · σ (Expression 6)
From the G image,
T3 = P_d(G)-{P_d(G) -P_hL(G)} · σ (Expression 7)
T4 = P_d(G)-{P_d(G) -P_hh(G)} · σ (Expression 8)
Get. Here, σ is a predetermined experimentally determined parameter, and σ is normally set to 3.0 to 3.5.
[0072]
(Procedure 5): Based on the threshold values T1 to T4 calculated in (Procedure 1) to (Procedure 4), a binary image is created in the image space according to the following (Equation 9). The background area (BG) is
{T1 ≦ R (i, j) <T2} ∩ {T3 ≦ G (i, j) <T4} (Equation 9)
Is obtained. Here, ∩ represents a logical product.
[0073]
(Procedure 6): The binary image obtained in (Procedure 5) is corrected and shaped.
[0074]
(Procedure 7): Labeling is performed as a target region where target particles exist in regions independent of each other other than the background region obtained in (Procedure 5).
[0075]
(Procedure 8): A feature amount is obtained for the pattern of the target area labeled in (Procedure 7), and the pattern of the target area is identified. In addition, a known feature amount is obtained by a well-known technique, and the pattern of each target region is identified, but the target pattern extracted for each color tone is determined based on the degree of staining of the target particle and the position state of multiple target particles (contact If the pattern obtained in (Procedure 6) is synthesized according to (the degree of superposition), a more stable target pattern can be obtained.
[0076]
The processing steps in this embodiment are summarized as follows.
[0077]
(1) Density correction processing: removes density unevenness derived from distortion of the optical system in the G image and the R image.
[0078]
(2) Region division: A G image and an R image are divided into a background region and a target region using a plurality of thresholds, and a binary image is created with the background region set to 0 and the target region set to 1.
[0079]
(3) Correction processing: Correction and shaping of a binary image, such as filling a target area and removing noise from a background area, are performed.
[0080]
(4) Labeling: A labeling process is performed for each connected component of a binary image, and a plurality of objects in the image are numbered.
[0081]
(5) Feature amount calculation: For each numbered object, a feature amount such as area, perimeter, average density, etc. is obtained.
[0082]
(6) Identification: Based on the obtained feature quantity, it is determined and classified what kind of urine sediment component each object is.
[0083]
In (3) to (6), known conventional techniques can be used including filter processing such as thickening processing and thinning processing.
[0084]
As a result of applying this example to 99 images obtained from four actual samples, the target particles could be extracted by dividing the region into 93 images out of 99 images. The six images that could not be extracted were images in which the target particles were unstained, images due to poor adjustment of the focal position, and the like.
[0085]
  FIG. 6 and FIG. 7 show the results of applying this example to actual samples (the processing of (3) above). 6 and 7, the horizontal (x) axis and the vertical (y) axis are each composed of 512 pixels of 0 to 511. FIG. 6 shows an example of a region division result of an image of a stained cell with good staining of the target particle, and FIG. 7 shows an example of a region division result of an image of a cell that is hardly stained with the target particle. According to the present embodiment, as in the example shown in FIG. 7, when the target particles are not sufficiently stained, for example, when there is no difference between the cytoplasm concentration and the background concentration and extraction is not complete.ButHowever, such particles can be extracted by the methods according to Example 2, Example 3, and Example 4 described below.
[0086]
(Example 2)
Hereinafter, for the sake of simplicity, cells (particles) that are hardly stained with a staining solution are abbreviated as unstained cells (particles), and cells that are stained well (particles) are abbreviated as stained cells (particles). Note that cells (particles) that are lightly stained but insufficiently stained and poorly stained are hereinafter referred to as unstained cells (particles).
[0087]
In the case where the target target particles are stained particles that are satisfactorily dyed with a staining agent, the region can be divided satisfactorily by the region dividing method of Example 1 described above.
[0088]
However, in fact, cells with high activity have poor staining properties, and there are cases where cells that are hardly stained (unstained cells) exist. Such unstained cells have been found to be frequently present in urine samples. In addition, there are particles in the urine sample that have poor staining properties and insufficient staining.
[0089]
In an image including unstained cells, the color tone of the target area and the background area are substantially the same, and there are many portions having the same or lower density as the background area in the target area. When unstained cells are present in the image, the region dividing method of Example 1 can only perform region division as shown in FIG. 7, for example, as described above, and the region where the cells exist is the same. Not connected by pixels with values. As shown in FIG. 7, when an accurate binary image cannot be obtained, the feature amount of the target region cannot be obtained accurately, and the possibility of erroneous classification becomes very high. In order to improve the identification accuracy of sediment components, it is necessary to divide the region accurately to identify cells and other particles that are poorly stained and hardly stained.
[0090]
Unstained cells (cells that are hardly stained with the staining solution) are almost transparent, and the area where the cells are present (target area) has almost the same color as the background area. It contains many density regions of almost the same level. First, the difference between an image containing unstained cells and an image containing stained cells will be described.
[0091]
8 and 9 show examples of density values of images including stained cells and unstained cells, respectively.
[0092]
8 to 13 shown below, the horizontal axis indicates the position of 512 pixels from 0 to 511 in the y direction of the image.
[0093]
FIG. 8 shows a plot of concentration values for one line (x = 300) in the y direction passing through a region containing stained cells in an image containing stained cells (cells that are well stained). FIG. 9 shows a plot of concentration values for one line (x = 340) in the y direction passing through a region containing unstained cells in an image containing unstained cells.
[0094]
8 and 9, the horizontal axis represents the position in the image (y coordinate value), and the vertical axis represents the density value. As the density value is closer to 0, the color is darker, and as the density value is closer to 255, the color is lighter. In FIG. 8, the G image shows that the density value of the stained cell region is lower than that of the background region. In this case, as shown in FIG. 8, it can be seen that, when a threshold value of, for example, −180 and −220 is set as the density value, the background region and the target region can be clearly divided in the G image (FIG. 8). For reference, the threshold value of ~ 190 is also shown). From a comparison between FIG. 9 and an actual image (not shown), it has been found that cells exist in a portion where the density change is large compared to the surroundings (in the vicinity of y = 400). From FIG. 9, the unstained cell region (target region) includes a portion having a large density value and a small portion each having a density value comparable to that of the background region, compared to the background region. As shown in FIG. 8, as shown in FIG. 8, it is indicated that accurate area division cannot be performed only by setting the same threshold value in the density value as shown in FIG. The same threshold is shown in the figure).
[0095]
Furthermore, it can be seen from FIG. 9 that in the image including unstained cells, the variation in the density value in the background region is small, but the variation in the density value in the region (target region) of the unstained cell is large.
[0096]
In the present embodiment, paying attention to the fluctuation of the density value, the magnitude of the density value change in the image is calculated, a threshold value is provided for each of the density value and the density value change, and the region division is performed. Then, the image is accurately segmented including unstained cells (cells that are hardly stained with the staining solution).
[0097]
First, as an index representing the magnitude of the change in density value, the distribution of density values in a small local area is considered. The density value at (x, y) on the image is defined as P (x, y), the local variance is defined as q (x, y), and q (x, y) is defined by the following (Equation 10).
[0098]
q (x, y) = {1 / (2n + 1)²} ΣΣ {P (i, j)}²-{1 / (2n + 1)^Four} {ΣΣP (i, j)}² ... (10)
In (Equation 10), addition Σ is performed for i = x−n to i = x + n and j = y−n to j = y + n.
[0099]
FIG. 10 and FIG. 11 show examples of dispersion values of images including stained cells and unstained cells, respectively (example of q (x, y) when n = 2). 10 and 11 are plots of the variance value q (x, y) at the same position (one line) as in FIGS. 8 and 9, respectively. The horizontal axis is the y-coordinate value of the image, and the vertical axis is (number 10 is a variance value calculated by 10).
[0100]
10 and 11, the large peak around y = 490 on the horizontal axis is due to the discontinuous component at the edge of the image and is irrelevant to the recognition target (cell) in the image. FIG. 11 shows a variance value q (x, y) of an image including unstained cells. Compared with FIG. 8, the variance value of the background region is close to 0, and the variance value of the region where cells are present is a large value. It has become. Therefore, the variance q is the threshold q_thSet q_thA region having a larger variance value is regarded as a target region of interest, and an image including unstained cells is divided into regions. FIG. 11 shows the threshold q_thIs shown as ~ 200.
[0101]
FIG. 10 shows a variance value q (x, y) of an image including stained cells. Compared with FIG. 8, the edge portion (horizontal axis, y = 200 and around y = 300) of the region where the cells exist is shown. The variance value shows a large value. However, since the inside of the cell is dyed with a uniform color tone, the variance value is small, and if the threshold value is set for the variance value and the region is divided, only the edge portion of the cell is divided as the target region. (In FIG. 10, as in FIG. 11, the threshold q_thIs shown as ~ 200. ). Therefore, the threshold value is set for the variance value, and the region is divided, and the threshold value is set for the concentration value, the region is divided, and the result of both the region divisions is overlapped, thereby dividing the region of both stained cells and unstained cells. Can be performed accurately.
[0102]
In the method described above, the local dispersion value of the density defined by (Equation 10) is used to detect the magnitude of density fluctuation in the image. However, since the apparatus configuration is simpler, it is more convenient. As a method for detecting the magnitude of density fluctuation, the density difference value defined by the following (Equation 11) is used. The density value at (x, y) is P (x, y), and the density difference value is r (x, y).
[0103]
r (x, y) = ΣP (x, y + i) −ΣP (x, y−i) (Equation 11)
In (Expression 11), addition Σ is performed for i = 1 to i = n. The calculation of (Equation 11) can be executed only by addition and subtraction, and the density value used to calculate the difference value at one point of the image is the density value of the pixels arranged one-dimensionally on the image. Therefore, high-speed processing synchronized with image signal transfer can be executed with a simple apparatus configuration.
[0104]
FIGS. 12 and 13 show examples of density difference values (density difference values when n = 2) of images including stained cells and unstained cells, respectively. 12 and 13 plot the density difference values at the same position (one line) as in FIGS. 8 and 9, respectively. As in the case of the variance value, the peak value near the horizontal axis 490 is due to the discontinuous portion at the image end.
[0105]
FIG. 12 shows a difference value of an image including stained cells. Compared with FIG. 8, the difference value takes a value that greatly changes in positive and negative at the edge portion of the target region, and the inside of the target region and the background region. Then, the difference value takes a value close to zero.
[0106]
FIG. 13 shows a difference value of an image including unstained cells. As can be seen from comparison with FIG. 9, the difference value takes a value close to 0 in the background region, and the difference value is positive or negative in the target (cell) region. The value is changing greatly.
[0107]
However, since there are many portions where the difference value takes a value close to 0 even within the target region, the difference value has two positive and negative thresholds, r_th1, R_th2(R_th1<0 <r_th2) And set r_th1<R (x, y) <r_th2When region division is performed with the region to be a background region, a region having a difference value close to 0 within the target region is divided as a background region. Further, when this method is applied to an image including stained cells, only the edge portion of the cell is divided as a target region. In FIG. 12 and FIG._th1~ -20, r_th2Shows the case of ~ 20.
[0108]
Here, consider the relationship between the density value and the density difference value. The point where the density difference value is 0 is a point where the density value takes a maximum value or a minimum value, and the density value at such a point is relatively larger or smaller than the average density value of the background region. It is likely to take a value. That is, as a result of performing region division by setting a threshold value for the density difference value, a portion that cannot be accurately divided (a portion inside the target region that is divided as a background region) is set with a threshold value for the region. If the division is performed, the target area can be divided. Therefore, by setting a threshold value for the difference value and performing region division, by setting a threshold value for the density value and performing region division, the results of both region divisions are superimposed (logical OR) to include unstained cells. Both an image and an image including stained cells can be accurately segmented.
[0109]
A result of applying the method described above to an image actually captured by the flow method will be described. When 51 examples of images containing squamous epithelial cells that were most likely to be unstained were examined, 8 images containing unstained particles appeared. In the method of Example 1, as a result of applying the above-described method to the images of 8 examples including unstained particles that were difficult to accurately divide into regions, the processing results obtained in the case of an image including stained particles; Equivalent results could be obtained. Hereinafter, among these images, an image including typical stained particles and an image including unstained particles will be described as examples.
[0110]
FIGS. 14 and 15 show the results of region segmentation of images containing stained cells and unstained cells, respectively, using a binarization method using local dispersion of concentrations (binary images (region segmented images)). An image in which correction processing is performed on a binary image)) is shown.
[0111]
14 to 17, the horizontal (x) axis and the vertical (y) axis are each composed of 512 pixels of 0 to 511.
[0112]
Using the density value of the G image, the variance value calculated as n = 2 in (Equation 10) and the density value of the G image were used in the region division processing. Specifically, if the density value at (x, y) is P (x, y), the variance value is q (x, y), and the binarization result is s (x, y),
P_th1<P (x, y) <P_th2 and q (x, y) <q_th:
s (x, y) = 0 (background: white)
otherwise: s (x, y) = 1 (target area: black) (Equation 12)
After obtaining a binary image represented by s (x, y), as a correction process (background area noise removal process, target area hole filling process), the target area is thickened once to be thinned. Was performed twice, and then a thickening process was further performed once (each of these processes is a well-known technique).
[0113]
Threshold value P set for density value_th1,P_th2Is the density value P giving the maximum value of the density histogram_m, Density value giving half the maximum value, P_s1,P_s2(P_s1<P_m<P_s2)
P_th1= P_m-3.5 (P_m-P_s1)
P_th2= P_m+3.5 (P_s2-P_m(...)
It was. The threshold value set for the variance value is q_th= 30.
[0114]
From FIG. 15, it is apparent that the result of region segmentation of the image including unstained cells is greatly improved compared to the result of Example 1 shown in FIG.
[0115]
Further, in the image including the stained cells (FIG. 14), the bacterium is present at the two positions (50, 51) on the right side of the cell region. It is detected at one place (52) on the right side. These bacteria were bacteria that could not be detected in the results of the method of Example 1 (FIGS. 6 and 7).
[0116]
Bacteria are present in two places (60, 61) on the left side of the cell area of the image (FIG. 14) containing the stained cells and in one place (62) on the left side of the cell area of the image (FIG. 15) containing the unstained cells. There is a smaller point, but this seems to be a noise component in the background area.
[0117]
FIGS. 16 and 17 show results of region segmentation of images including stained cells and unstained cells, respectively, using a binarization method using density difference values (binary image (binary segmented binaries)). An image in which correction processing is performed on the image)) is shown. Using the density value of the G image, the difference value calculated as n = 2 in (Equation 11) and the density value of the G image were used in the region division processing. Specifically, if the density value at (x, y) is P (x, y), the difference value is r (x, y), and the binarization result is s (x, y),
P_th1<P (x, y) <P_th and r_th1<R (x, y) <r_th2:
s (x, y) = 0 (background: white)
otherwise: s (x, y) = 1 (target area: black) (Expression 14)
Then, after obtaining a binary image represented by s (x, y), correction processing was performed.
[0118]
Correction processing and density threshold P_th1, P_th2The same method as in the case of area division using the binarization method using the local dispersion of the density was used. The threshold value set for the difference value is | r_th1│, │r_th2│ can be set to about 15-25, r_th1= -20, r_th2= 20.
[0119]
FIG. 17 clearly shows that the result of segmentation of an image including unstained cells is greatly improved as compared to the result of Example 1 shown in FIG.
[0120]
In addition, in the image including the stained cells (FIG. 16), the bacteria are present at the two locations (50, 51) on the right side of the cell region. It is detected at one place (52) on the right side. These bacteria are those that could not be detected in the results of the method of Example 1 (FIGS. 6 and 7).
[0121]
When comparing FIG. 7, FIG. 15 and FIG. 17, compared with the region dividing method of the first embodiment, two types of region dividing methods using the above-described local distribution of density and binarization method using difference values are used. It can be seen that this is effective for the region division of unstained cells. Of the two types of methods, the method using the dispersion value of concentration is more effective for the region division of unstained cells. However, the method using the difference value of the density can simply configure the apparatus. Therefore, details of the region division method using the difference value will be described.
[0122]
The results shown in FIG. 16 and FIG.Calculation (This is the result when n = 2 in Equation 11). The number of neighboring pixels used when calculating a difference value of a certain pixel is represented by 2n + 1, and this number is referred to as a mask size here. For example, when n = 2, the calculation of the difference value of a certain pixel requires a total of 5 pixels, centering on that pixel, 2 pixels before and after in the y direction, and the mask size = 5. When the mask size = 1, it means that the difference process is not performed.
[0123]
FIG. 18 shows the result of area division by changing the mask size. FIG. 18 shows an image made up of 512 × 512 pixels. In order to investigate the influence of the mask size on the area division, the value of n was changed and the area division was performed. The example of FIG. 18 is a result of region segmentation of an image including the same unstained cells as in FIGS. 7, 15, and 17. The result of mask size = 5 in FIG. 8 corresponds to the result of FIG. In the example of FIG. 18, the result with mask size = 9 gives the best region segmentation result (however, a white region indicating an erroneously identified background is present in a part inside the region where cells exist. Mixed).
[0124]
Moreover, the cell 52 is detected also in FIG. 18 similarly to the result of FIG. 15, FIG. Note that the result when the difference processing is not performed with the mask size = 1 is similar to the result according to the first embodiment shown in FIG.
[0125]
Next, let us consider the influence of the mask size on the frequency axis.
[0126]
FIG. 19 shows the result of discrete Fourier transform of the signal component (density value) of one line (512 pixels) of the G image in the y direction of the background area as a dotted line, and the y direction passing through the unstained cell area as a solid line. The result of discrete Fourier transform of the signal component (density value) of the G image of one line (512 pixels) is shown. The horizontal axis represents the spatial frequency expressed as a unit length of one side (512 pixels) of the image, and the vertical axis represents the signal power when the maximum density value is normalized to 0.1.
[0127]
From FIG. 19, since the signal in the background area is a noise component, it is distributed with almost the same power except for the DC component, and the image signal of the line passing through the unstained cell area is the background area. It can be seen that the signal components of frequencies 1 to 150 are stronger than
[0128]
If (Equation 11) used to emphasize the region of unstained cells is regarded as a digital filter, the transfer function uses the z-transform,
H (z) = Σ {z ↑ (k) −z ↑ (−k)} (Equation 15)
It is expressed. Here, z ↑ (y) indicates the power (y) of z (y = k, y = −k). In (Expression 15), addition Σ is performed for k = 1 to k = n. The frequency response at this time is

It is expressed. Ω represents a spatial frequency whose unit length is two pixels on the image, and j is an imaginary unit.
[0129]
FIG. 20 shows the frequency amplitude characteristics | H (Ω) |) when n = 1, 2, and 3.
[0130]
The horizontal axis in FIG. 20 represents the spatial frequency expressed as the unit length of the length of one side (512 pixels) of the image, and the vertical axis represents the value obtained by normalizing the filter response using the mask size. . In addition, the phase characteristic is constant at π / 2, and the effect of this filter is to emphasize the frequency component peculiar to the image containing unstained cells and to separate it from the background, and by shifting the phase by 90 degrees, And a portion that becomes 0 when binarized and a portion that becomes 0 when binarized using the difference value.
[0131]
From FIG. 20, in the case of n = 2 (mask size 5) and n = 3 (mask size 7), the signal components in the medium frequency region to the low frequency region peculiar to the image including unstained cells are strengthened. When the mask size becomes excessive, although not shown in FIG. 20, the value of n is increased as in the case of n = 5.TheAs the process proceeds, the low frequency region of the signal components peculiar to the image including unstained cells is weakened. For this reason, the value of n is suitably about 2 to 4 (mask size 5 to 9).
[0132]
FIG. 21A and FIG. 21B show the results of examining the influence of the mask size on the shape of the region divided as the target region, using stained red blood cell images. In FIGS. 21A and 21B, the unit of length is the number of pixels.
[0133]
Among the objects to be classified, red blood cells are the second smallest after bacteria, and the shape is almost circular. Therefore, the red blood cell image is divided into regions by changing the mask size, and changes in the shape of the extracted target regions are evaluated. Thus, it is possible to evaluate deterioration in the image processing process (difference between the true shape of the target particle and the shape of the target region divided into regions).
[0134]
FIG. 21A is a plot of the area and perimeter of the target region (red blood cells) obtained for various mask sizes (shown as numbers in the figure). As the mask size increases, both the area and the perimeter increase, and the target area expands.
[0135]
FIG. 21B is a plot of the projection lengths in the x and y directions obtained when the red blood cells divided into regions are projected onto the x axis and the y axis, respectively. Since erythrocytes are almost circular, the projection lengths on the x-axis and y-axis are almost the same value when the difference processing is not used (mask size = 1). However, when the difference process is performed, the projection length in the y-axis direction, which is the direction of the difference process, is larger than the projection length in the x-axis direction, and the target area is a vertically long ellipse. This is because, as can be seen from FIG. 20, the high-frequency component is suppressed and the target edge is expanded by the difference processing, and is thus expanded in the processing direction.
[0136]
In addition, this tendency becomes stronger as the mask size increases. As described above, it is necessary to determine the mask size in consideration of two points of improvement in binarization of an image including unstained cells and deterioration of a binary image of an image including stained cells.
[0137]
Improvement of binarization of an image containing unstained cells enables more accurate calculation of feature parameters and leads to an improvement in recognition rate of pattern recognition at a later stage, but deterioration of the binary image of an image containing stained cells. Makes the feature parameters inaccurate and affects the classification accuracy, especially for small objects. Further, the obtained binary image changes depending on the threshold value used for binarization. In general, the smaller the absolute value of the threshold, the better the region division of the unstained cells, but the easier it is to pick up noise components in the background region. It is necessary to set an optimum threshold value while paying attention to this point.
[0138]
FIG. 22 shows a flowchart of processing for explaining the procedure for dividing the region of the particle image of this embodiment. As described above, the particle image to be processed is stored in the image memory. In this embodiment, the image to be used is a G image and is represented by a density value P (x, y) at the position (x, y) of each pixel.
[0139]
Hereinafter, the process will be described according to the process steps shown in FIG.
[0140]
(Step S1): First, the density value P (x, y) at the pixel (x, y) is read from the image memory.
[0141]
(Step S2): The density value P (x, y) is converted into a variable u (x, y) representing the binary image shown below.
[0142]
When T2 '<P (x, y) <T1': u (x, y) = 0 (Equation 17)
otherwise: u (x, y) = 1 (Expression 18)
Here, T1 'and T2' represent threshold values. For example, T1 'and T2' are determined as follows. First, the density value P (x, y) of the image stored in the image memory is read, a histogram of density values is created, and density threshold values T1 'and T2' are determined from the shape of the histogram. Usually, since the area of the background portion is the largest on the image, the average density of the background takes the maximum value on the histogram. As shown in FIG. 5, the density value P giving the maximum value of this histogram._dIs detected, and this maximum value is defined as Pmax. Next, the density value giving the value of Pmax / 2 is detected, and P_hL, P_hh(P_hL<P_d<P_hh).
[0143]
An appropriate constant k is determined, and threshold values T1 'and T2' are determined from (Equation 19) and (20).
[0144]
T1 '= P_d+ K (P_hL-P_d(Equation 19)
T2 '= P_d+ K (P_hh-P_d) (Equation 20)
The constant k is obtained experimentally in advance.
[0145]
As shown in (Expression 17), when T2 ′ <P (x, y) <T1 ′, the pixel (x, y) is considered to be in the background area ((Expression 17)), and T2 ′ ≧ P (x , Y) or P (x, y) ≧ T1 ′, the pixel (x, y) is a target region (a region other than the background region, which indicates a region where particles to be distinguished from the background region exist). ((Equation 18)).
[0146]
(Step S3): The density values of a plurality of pixels near the pixel (x, y) are read from the memory.
[0147]
(Step S4): The magnitude of the density value change (the magnitude of the density value change at the pixel (x, y) is q (x, y)) is any of the following (a) to (d). Use one of these methods to calculate.
[0148]
(A); As shown in (Expression 21) or (Expression 22), the density values of a plurality of pixels in the vicinity of the pixel (x, y) read out in step S3 are two regions adjacent to each other on the image. The sum of the density values of the pixels belonging to the respective areas is obtained, and then the difference between the two obtained sums ((Equation 21)) or its absolute value ((Equation 22)) Ask for.
[0149]
q (x, y) = Σ {P (x, y + i) −P (x, y−i)} (Expression 21)
q (x, y) = | Σ (P (x, y + i) −P (x, y−i)) | (Equation 22)
For example, the density values of four pixels near the pixel (x, y) read out in step S3, P (x, y-2), P (x, y-1), P (x, y + 1), P From (x, y + 2), q (x, y) is calculated by (Equation 21) or (Equation 22). In (Expression 21) and (Expression 22), addition Σ is performed for i = 1 to i = n.
[0150]
(B): The variance ((Equation 23)) or standard deviation ((Equation 24)) of the plurality of pixels in the vicinity of the pixel (x, y) read out in step S3 is obtained, and the density value change Let q (x, y) be the magnitude of.
[0151]
q (x, y) = {1 / (2n + 1)²} ΣΣ {P (x + i, y + j)}²-{1 / (2n + 1)²}²{ΣΣP (x + i, y + j)}²... (Equation 23)
q (x, y) = √ <{1 / (2n + 1)²} ΣΣ {P (x + i, y + j)}² -{1 / (2n + 1)²}²{ΣΣP (x + i, y + j)}²> (Equation 24)
Note that √ <> indicates the square root in <>. In (Expression 23) and (Expression 24), addition Σ is performed for i = −n to i = n and j = −n to j = n, respectively.
[0152]
For example, in step S3, the density values P (i, j) of nine pixels satisfying x-1 ≦ i ≦ x + 1 and y−1 ≦ j ≦ y + 1 are read out, and (Equation 23) or (Equation 24) , Q (x, y).
[0153]
(C); obtained by taking the product of the density value of a plurality of pixels in the vicinity of the pixel (x, y) obtained in step S3 and the weight value w corresponding to the geometric arrangement of each pixel. The sum of the products is defined as the magnitude q (x, y) ((Equation 25)) of the density value change.
[0154]
q (x, y) = {ΣΣw (i, j) P (x + i, y + j)} (Equation 25)
In (Equation 25), addition Σ is performed for i = −n to i = n and j = −n to j = n, respectively. For example, nine weight values w (k, m) (k, m = −1, 0, 1) are set in advance, and the density values P (i, j) (x−) of the nine pixels are set in step S3. 1 ≦ i ≦ x + 1, y−1 ≦ j ≦ y + 1) is read, and q (x, y) is calculated by (Equation 25). The weight values are, for example, w (0,0) = 4, w (-1,0) = w (1,0) = w (0, -1) = w (0,1) = 1, w (− 1, -1) = w (-1,1) = w (1, -1) = w (1,1) = 0.
[0155]
(D): After obtaining the magnitude q (x, y) of the density value change using any one of the above methods (a) to (c), the density value change obtained Considering the magnitude as a density value P (x, y),
P (x, y) = q (x, y) (Equation 26)
The density value change magnitude q (x, y) may be calculated by any one of the methods (a) to (c).
[0156]
For example, after calculating the value of q (x, y) for all the pixels (x, y) by the method of (a) above, applying (Equation 26), and then the method of (b) above Q (x, y) is calculated again using.
[0157]
(Step S5): The magnitude q (x, y) of change in density value is converted into a variable v (x, y) representing the binary image shown below.
[0158]
When q (x, y) takes only a positive value, the threshold T3 'is used,
When q (x, y) <T3 ′: v (x, y) = 0 (Equation 27)
When q (x, y) ≧ T3 ′: v (x, y) = 1 (Equation 28)
When q (x, y) <T3 ′, the pixel (x, y) is considered to be in the background area (Equation 27), and when q (x, y) ≧ T3 ′, the pixel (x, y) y) is considered to be in the target area ((Equation 28)). The threshold value T3 'can be obtained experimentally in advance.
[0159]
When q (x, y) takes a negative value (step S5 '), the threshold value T3 "and the threshold value T4' are used,
When T4 ′ <q (x, y) <T3 ″: v (x, y) = 0 (Equation 29)
otherwise: v (x, y) = 1 (Expression 30)
When T4 ′ <q (x, y) <T3 ″, the pixel (x, y) is considered to be in the background area (Equation 29), and T4 ′ ≧ q (x, y) or q ( When x, y) ≧ T3 ″, the pixel (x, y) is considered to be in the target area ((Expression 30)). The threshold values T3 ″ and T4 ′ can be experimentally obtained in advance.
[0160]
The selection of S5 and S5 'in (Step S4') is automatically determined depending on how q (x, y) is defined in (Step S4).
[0161]
(Step S6): The logical sum of u obtained in (Step S2) and v obtained in (Step S5) or (Step S5 ′) is obtained as a result of binarization at the pixel (x, y). To do.
[0162]
(Steps S7, S8): If the pixel (x, y) is the last pixel point to be processed, the process proceeds to step S9; otherwise, the pixel (x, y) is the next pixel point. To repeat the process of steps S1 to S6 (steps S7 and S8).
[0163]
(Step S9): A correction process is performed on the binary image obtained by the above processing with the target area set to 1 and the background area set to 0, such as filling the target area and removing noise from the background area.
[0164]
(Step S10): For each connected component of the binary image, a labeling process by a well-known technique is performed, and a plurality of objects in the image are numbered.
[0165]
(Step S11): For each numbered object, a feature amount by a known technique such as area, perimeter, average density, projection length, etc. is obtained.
[0166]
(Step S12): Using the obtained feature amount, each object is classified and identified as any kind of urine sediment component.
[0167]
FIG. 23 shows an apparatus configuration for carrying out this embodiment. Particles in the liquid are irradiated by a pulse lamp 101, magnified by an optical magnification system (not shown), photographed by a television camera 102, and converted into an analog electrical signal. The output of the television camera 102 is converted into a digital signal by the A / D converter 103 and then stored in the image memory 104 via the data bus 110. The threshold value calculation unit 105 reads out the density value of the image stored in the image memory 104 via the data bus 110 and creates a histogram of the density value. According to (Equation 19) and (Equation 20), an area based on the histogram shape is obtained. Density threshold values T1 ′ and T2 ′ used in the dividing unit 106 are determined.
[0168]
The area dividing unit 106 divides the area using the density value obtained from the image memory 104 and the threshold values T1 ′ and T2 ′ obtained from the threshold calculating unit 105 via the data bus 110, and sets the target area as 1 After binarization with the background area set to 0, the binarization result is stored in the image memory 107 via the data bus 110. All of the above processing is controlled by the computer 108 via the CPU bus 109.
[0169]
FIG. 24 shows a configuration of the area dividing unit 106 in FIG. The input unit 200 is connected to the data bus 110 in FIG. 23 and controls data input via the data bus 110. In the density binarization circuit 300, first, the threshold values T1 ′ and T2 ′ are read from the threshold value calculation unit 105 (FIG. 23) via the input unit 200, and the values of T1 ′ and T2 ′ are stored in the threshold value memory 310. To do. Next, the density value of one pixel on the image is read from the image memory 104 via the input unit 200, and the density value is compared with the threshold values T1 'and T2' by the comparator 320. Assuming that one point on the image is (x, y) and the density value at that point is P (x, y), the comparator 320 compares the density value P (x, y) with the threshold values T1 ′ and T2 ′. , T2 ′ <P (x, y) <T1 ′, the pixel is regarded as a background area and 0 is output, otherwise it is regarded as a target area and 1 is output. .
[0170]
In parallel with the above binarization of density, the density change binarization circuit 400 operates. In the density change binarization circuit 400, first, density values of a plurality of neighboring pixels on the image (x, y) are read from the image memory 104 (FIG. 23) via the input unit 200 and stored in the memory 410. The density change calculation unit 420 uses the density values of the (x, y) neighboring pixels held in the memory 410 to change the density value using any one of the following methods (a) to (d). Calculate the size. Let q (x, y) be the magnitude of the change in density value at (x, y).
[0171]
(A) The density value of the pixel stored in the memory 410 is divided into pixels belonging to two adjacent areas on the image, and the sum of the density values of the pixels belonging to each area is obtained. Next, the difference between the two obtained sums or the absolute value thereof is obtained, and the magnitude of change in the density value is defined as q (x, y). For example, the density values P (x, y−2), P (x, y−1), P (x, y + 1), and P (x, y + 2) of four pixels are stored in the memory 410, Q (x, y) is calculated by (Equation 31) or (Equation 32).
[0172]
q (x, y) = ΣP (x, y + i) −ΣP (x, y−i) (Equation 31)
q (x, y) = | ΣP (x, y + i) −ΣP (x, y−i) |
In (Expression 31) and (Expression 32), addition Σ is performed for i = 1 to i = 2, respectively. FIG. 25 shows a specific configuration example of the density change binarization circuit 400 in this case. The memory 410 is composed of four memories 411 to 414 each storing one density value. Concentration values P (x, y−2), P (x, y−1), P (x, y + 1), and P (x, y + 2) are stored in the memories 411 to 414, respectively.
[0173]
The density change calculation unit 420 includes two adders 421 and 422 and one subtracter 423, and performs the calculation represented by (Equation 31). The output of the adder 421 outputs the value calculated by the second term on the right side of (Equation 31), and the adder 422 outputs the value calculated by the first term on the right side of (Equation 31). The output of the subtracter 423 outputs a value calculated on the right side of (Equation 31). When the subtracter 423 has a function of outputting an absolute value, the calculation shown in (Expression 32) can be performed.
[0174]
(B) The variance or standard deviation of the density values stored in the memory 410 is obtained and set as the magnitude q (x, y) of the density value change. For example, the density values P (i, j) of nine pixels that satisfy x−1 ≦ i ≦ x + 1 and y−1 ≦ j ≦ y + 1 are stored in the memory 410, and the following (Expression 33) or ( Q (x, y) is calculated by Equation 34).
[0175]
q (x, y) = {1/9} ΣΣ {P (x + i, y + j)}²− {1/81} {ΣΣP (x + i, y + j)}² ... (Expression 33)
q (x, y) = √ {{1/9} ΣΣ {P (x + i, y + j)}²− {1/81} {ΣΣP (x + i, y + j)}²> (Formula 34)
Note that √ <> indicates the square root in <>. In (Expression 33) and (Expression 34), addition Σ is performed for i = −1 to i = 1 and j = −1 to j = 1, respectively.
[0176]
FIG. 26 shows a specific circuit configuration example of the memory 410 and the density change calculation unit 420 in this case. The memory 410 is composed of nine memories 811 to 819 each storing one density value. The memories 811 to 819 store the density values P (i, j) of nine pixels that satisfy x−1 ≦ i ≦ x + 1 and y−1 ≦ j ≦ y + 1.
[0177]
The density change calculator 420 includes a square calculator 821 including nine square calculators, adders 822 and 824, dividers 823 and 825, a square calculator 826, and a subtractor 827. The values read from the memories 811 to 819 are each divided into two, one being input to the square calculator 821 and the other being input to the adder 824. The adder 822 calculates the sum of the outputs of the nine square calculators of the square calculator 821, and the divider 823 divides the output value of the adder 822 by 9. The divider 823 outputs the value calculated by the first term on the right side of (Expression 33).
[0178]
The adder 824 calculates the sum of the nine density values read from the memories 811 to 819, and the divider 825 divides the output value of the adder 824 by 9. The square calculator 826 outputs the value calculated by the second term on the right side of (Expression 33). The subtractor 827 calculates the difference between the output of the divider 823 and the output of the square calculator 826. The subtracter 827 outputs the value calculated on the right side of (Equation 33). If a circuit for calculating the square root is provided after the subtractor 827, the value calculated on the right side of (Equation 34) can be output.
[0179]
(C) The product of the density value of the pixel stored in the memory 410 and the weight value w corresponding to the geometric position of the pixel image is taken, and the total sum of the obtained products is the magnitude of the change in the density value. Let q (x, y). For example, nine weight values w (k, m) (k, m = −1, 0, 1) are set in advance, and the density values P (i, j) (x −1 ≦ i ≦ x + 1, y−1 ≦ j ≦ y + 1) is held, and q (x, y) is calculated by the following (Equation 35).
[0180]
q (x, y) = {ΣΣw (i, j) P (x + i, y + j)} (35)
The weight values are, for example, w (0,0) = − 4, w (−1,0) = w (1,0) = w (0, −1) = w (0,1) = 1, w ( −1, −1) = w (−1,1) = w (1, −1) = w (1,1) = 0. In (Expression 35), the addition Σ is performed for i = −1 to i = 1 and j = −1 to j = 1, respectively.
[0181]
FIG. 27 shows a specific circuit configuration example of the memory 410 and the density change calculation unit 420 in this case. The memory 410 is composed of nine memories 931 to 939 each storing one density value. In the memories 931 to 939, density values P (x-1, y-1), P (x, y-1), P (x + 1, y-1), P (x-1, y), P (x , Y), P (x + 1, y), P (x-1, y + 1), P (x, y + 1), and P (x + 1, y + 1) are stored. In the weight value memories 941 to 949, the weight values w (-1, -1), w (0, -1), w (1, -1), w (-1, 0), w (0, 0) are stored. , W (1, 0), w (-1, 1), w (0, 1), w (1, 1) are set in advance. The multipliers 951 to 959 calculate the product of the density value and the weight value, and the adder 961 obtains the sum of the nine obtained products.
[0182]
(D) Using any one of the methods (a) to (c) described above, the magnitude of the change in the density value is obtained, and then the magnitude of the obtained density value change is expressed as the density value. Therefore, the magnitude of the change in the density value may be calculated again using any one of the methods (a) to (c), and the obtained value may be set to q (x, y).
[0183]
For example, after calculating the value of q (x, y) for all pixels (for example, 512 × 512 pixels) by the method of (c), using this q (x, y),
P (x, y) = q (x, y) (36)
In addition, q (x, y) is calculated again using the method (b). In this case, the memory 410 and the density change calculation unit 420 are replaced with a circuit of the density change calculation unit having the configuration shown in FIG.
[0184]
First, the load sum calculation circuit 900 calculates the value of the right side of (Equation 35) for all pixels, and stores the obtained value in the image memory 1000. Next, the value stored in the image memory 1000 is calculated as a density value, and the value on the right side of (Expression 33) is calculated and output by the variance calculation circuit 800. The configurations of the load sum calculation circuit 900 and the variance calculation circuit 800 are the configurations shown in FIGS. 27 and 26, respectively.
[0185]
The comparison unit 430 in FIG. 24 includes a threshold memory 431 and a comparator 432 as shown in FIG. In the threshold memory 431, threshold values T3 ′, T3 ″, and T4 ′ are stored in advance, and when the magnitude q (x, y) of the density change takes only a positive value, q (x, y) <T3 If ', the pixel is regarded as a background region and 0 is output, otherwise it is regarded as the target region and 1 is output.
[0186]
When q (x, y) takes a negative value, 0 is output if T4 ′ <q (x, y) <T3 ″, and 1 is output otherwise. Threshold values T3 ′, T3 ″, T4 'Can be obtained experimentally in advance. Next, the logical sum circuit 500 takes a logical sum of the output of the density binarization circuit 300 and the output of the density change binarization circuit 400. The above processing is applied to all the pixels to obtain a binary image in which the target area is 1 and the background area is 0. The binary image obtained in this way is further filtered by the filter unit 600 to perform processing for filling the target region by thickening processing or the like, and for removing noise from the background region by thinning processing or the like. Correction is performed, and the corrected binary image is output from the output unit 700 to the data bus 110 (FIG. 24).
[0187]
The processing steps in this embodiment are summarized as follows.
[0188]
(1) Density correction processing: removes density unevenness derived from distortion of the optical system in the G image.
[0189]
(2) Region division: A binary image in which the G image is divided into a background region and a target region based on the density and the amount representing density change in the G image, and the background region is 0 and the target region is 1. create.
[0190]
(3) Correction processing: Correction and shaping of a binary image, such as filling a target area and removing noise from a background area, are performed.
[0191]
(4) Labeling: A labeling process is performed for each connected component of the binary image, and a plurality of objects in the image are numbered.
[0192]
(5) Feature amount calculation: For each numbered object, a feature amount such as area, perimeter, average density, etc. is obtained.
[0193]
(6) Identification: Based on the obtained feature quantity, it is determined and classified what kind of urine sediment component each object is.
[0194]
In (3) to (6), known conventional techniques can be used including filter processing such as thickening processing and thinning processing.
[0195]
(Example 3)
In this embodiment, instead of the configuration in which the threshold is set for the density of the G image in the configuration of the second embodiment and the region is divided, the configuration of the first embodiment (the threshold is set in the density of the G image and the R image is divided into regions The target particles are extracted more accurately using the above-described configuration.
[0196]
According to the results applied to a large number of samples, the region division method of Example 1 can extract the stained cells more accurately than the unstained cells. In this method, since only the G image is used, it was found that the extraction of the stained cells is inferior to that in Example 1 when the G image of the stained particle image is unclear. In this embodiment, the threshold value is provided for the density of the G image and the R image to divide the region. Therefore, even when the G image is unclear, the information of the R image is used, and either the first embodiment or the second embodiment is used. This makes it possible to extract target particles more accurately.
[0197]
The processing steps in this embodiment are summarized as follows.
[0198]
(1) Density correction processing: removes density unevenness derived from distortion of the optical system in the G image and the R image.
[0199]
(2) Region division: A G image and an R image are divided into a background region and a target region using a plurality of thresholds, and a binary image is created with the background region set to 0 and the target region set to 1.
[0200]
(3) Region division: A G image is divided into a background region and a target region based on an amount representing density change in the G image, and a binary image is created with the background region set to 0 and the target region set to 1.
[0201]
(4) Region division: Logical sum of the binary images obtained in (2) and (3).
[0202]
(5) Correction processing: Correction and shaping of a binary image such as hole filling of a target region and noise removal of a background region are performed.
[0203]
(6) Labeling: A labeling process is performed for each connected component of a binary image, and a plurality of objects in the image are numbered.
[0204]
(7) Feature amount calculation: A feature amount such as an area, a perimeter, an average density, etc. is obtained for each numbered target.
[0205]
(8) Identification: Based on the obtained feature quantity, it is determined and classified what kind of urinary sediment component each object is.
[0206]
In (5) to (6), known conventional techniques including filter processing such as thickening processing and thinning processing can be used. Further, the processing of (5) may be performed on the processing results of (2) and (3), and the processing of (5) may be omitted.
[0207]
It goes without saying that either of the processes (2) and (3) for dividing the area may be performed first or in parallel.
[0208]
From the red component image (R image), green component image (G image), and blue component image (B image) of the target particle, the red component image (R image) of the target particle in the same manner as in Example 1. ), A green component image (G image) is selected, and, similarly to the second embodiment, a green component image (G image) is selected, and the configuration of the first embodiment is partially included in the configuration of the second embodiment. As a result of applying to 51 examples of images including the flattened epithelial cells that are most likely to be unstained, as described in Example 2, a good region was obtained in all 51 images including 8 unstained particles. The division result was obtained. In other words, the configuration of Example 1 was incorporated in a part of the configuration of Example 2, and satisfactory region division that could not be obtained by the single method of Example 1 and Example 2 could be performed. As a result, almost all target particles, including unstained cells that were difficult to be divided as one region, can be divided as one connected target region by a single method. The accuracy of component identification could be improved.
[0209]
In this example, even when particles that are sufficiently stained and particles that are difficult to be dyed (particles that are not sufficiently stained, particles that are hardly stained) are mixed together, the background region is accurately identified, The region can be stably divided for each particle image, and a more accurate binary image can be obtained. As a result, the feature amount of the target region can be obtained more accurately, and the classification error of the target particle can be reduced. In this way, even when target particles with different degrees of staining are mixed, target particles that are poorly dyeable and hardly stained are included in the sample, accurate segmentation can be performed for each particle image, and various sediment components The identification accuracy can be improved.
[0210]
FIG. 30 and FIG. 31 show examples of processing results on actual samples according to this example.
[0211]
The sample in FIG. 30 is the same sample as the examples shown in FIGS. 6, 14, and 16, and shows an example of cells (stained cells) with good staining. The sample in FIG. 31 is the same sample as the example shown in FIGS. 7, 15, and 17, and shows an example of cells (stained cells that are not stained well).
[0212]
30 and 31, binary images 251 and 261 are binary images 252 and 262, respectively, which are obtained as a result of region segmentation by setting threshold values for the density values of the G image and the R image of the first embodiment. The results of region segmentation by the method of setting a threshold value for the amount representing the change in the density value of the G image in Example 2 (n = 2 and using the difference value) are shown. The binary image 253 shows the result of taking the logical sum of the binary images 251 and 252, and the binary image 263 shows the result of taking the logical sum of the binary images 261 and 262. It should be noted that, based on the result of region division of these binary images 251, 261, 252, 262, 253, and 263, correction processing (binary image correction and shaping processing such as target region filling and background region noise removal) is performed. Not done. The binary images 254 and 264 are images obtained by performing correction processing on the binary images 253 and 263, respectively. 6 is obtained when the binary image 251 is corrected, and the image of FIG. 7 is obtained when the binary image 261 is corrected.
[0213]
In the method of Example 3, even if it is difficult to ideally extract the particles in the sample by the single methods of Examples 1 and 2, it is possible to extract the particles in the image with almost no oversight. Become.
[0214]
Example 4
In this embodiment, any two binary images obtained by the region dividing method described in the first to third embodiments, or a logical operation (logical sum) of three binary images is newly added. A binary image is obtained, and target particles are extracted more accurately.
[0215]
FIG. 29 is a diagram schematically illustrating the synthesis of two binary images obtained by area division by two different methods (area division methods based on the first embodiment and the second embodiment) in the present embodiment. . A new binary image 230 is synthesized from the binary image 210 obtained from the region dividing method based on the first embodiment and the binary image 220 obtained from the amount representing the change in density value of the G image in the second embodiment. . This composition is obtained by the logical sum of the two binary images 210 and 220. In the particle image 212 of the binary image 210, only a part of the periphery of the particle image 222 obtained in the binary image 220 is extracted. Conversely, in the particle image 221 of the binary image 220, the binary image 210 is extracted. Only the part of the inside of the particle image 211 obtained in (the part of the inside of the particle image 211 is identified as the background region) is extracted.
[0216]
In the new binary image 230 synthesized by the logical sum of the two binary images 210 and 220, the particle images 231 and 232 are obtained, and the quality of the image is improved by compensating for the drawbacks of the respective region segmentation methods. A binary image is obtained, and target target particles can be extracted more accurately. In the above description, a logical sum is taken as the synthesis method, but it is appropriately set depending on the processing conditions (staining conditions, etc.) of the sample to be image-processed, the type of sample, the type of target particle, and the like.
[0217]
In each of the above-described embodiments, the determination of which image to select among red, green, and blue images as the image to be used for region segmentation depends on the spectral characteristics of the sample to be analyzed, the level and variations of the background signal, and the target In consideration of the signal level and variation in the region where the particles exist, the region division method described in each embodiment may be performed. Furthermore, a logical sum is used as a logical operation for performing region segmentation to extract the target particles of interest, but depending on the processing conditions (staining conditions, etc.) of the sample to be image processed, the type of sample, the type of target particle, etc. Needless to say, it is changed as appropriate.
[0218]
The urine sample is usually added with a staining solution, and the cells among the sediment components are usually stained in colors such as blue, red, and magenta. Since the background is white, simply focus on the color difference between the background and the sediment component, set threshold values for the image signals of the red image, green image, and blue image. In the conventional region dividing method in which the other region is the target region where the target particles are present, particles that cannot be sufficiently stained cannot be detected accurately. The reason is that, among the cells, cells that are not completely dead and highly active have the property of being difficult to absorb the stain and are transparent without being stained. Since the color tone is almost the same as that of the area, the conventional simple area division method cannot accurately extract the cell area.
[0219]
Cells that are not sufficiently stained are colorless and transparent, and light is refracted and scattered due to non-uniformity inside and on the cell surface, and the density value of the image of the area where the cells are present varies greatly. For this reason, in the simple area division method, an area where one cell exists is divided into a plurality of small areas, and each independent area is erroneously identified as a different sediment component, and sediment is generated. The accuracy of component identification is reduced.
[0220]
In the present invention, paying attention to such a small change depending on the position of the density value of the cell image, an amount representing the change in density at each point in the image is calculated, and at the same time as setting a threshold value for the density value, By providing a threshold for the amount representing the change and performing region division, accurate region division can be performed even when the target particles are not sufficiently stained.
[0221]
(Example 5)
An apparatus for segmenting a particle image to which the method described in the first to fourth embodiments is applied will be described below with reference to FIG.
[0222]
Image signals of G image, R image, and B image in a predetermined shooting area (FIG. 3) from the TV camera 102 are converted into digital image data related to, for example, 512 × 512 pixels by the A / D converter 103. Is done. Each process described below is executed for each pixel of the image data, and a region where particles are present and a background region are identified, and a binary image divided into regions is obtained.
[0223]
The density correction circuit 72 uses correction data stored in advance in the correction data memory 70 to correct density unevenness due to distortion of the imaging optical system, and is based on (Equation 2) or (Equation 3). Then, correction of density unevenness is executed. The correction result is stored in the image memory 90. The above configuration is a configuration common to apparatuses that execute the first to third embodiments.
[0224]
Hereinafter, an example of an apparatus configuration for executing the first to fourth embodiments will be described.
[0225]
(A) The apparatus for executing the first embodiment includes a histogram calculation circuit 74, a threshold value calculation circuit 76, a density binarization circuit 78, and a binary image memory 80 in the configuration shown in FIG. Yes. The histogram calculation circuit 74 obtains a histogram of the G image and the R image using the R image data and the G image data whose density unevenness has been corrected by the density correction circuit 72. The threshold calculation circuit 76 calculates thresholds T1, T2, T3, and T4 based on (Equation 5) to (Equation 8), respectively. The density binarization circuit 78 uses these threshold values T1, T2, T3, and T4 to detect a background area based on (Equation 9) from the G image and R image data, and binarize the image. The data is stored in the binary image memory 80.InStore. In the apparatus that executes the first embodiment, the density binarization circuit 78 and the binary image memory 80 also serve as the binarization circuit 96 and the binary image memory 98, respectively. (B) The configuration of the apparatus that executes the second embodiment detects the magnitude of the change in density between pixels in the R image data in which the density unevenness read out from the image memory 90 is corrected in the configuration shown in FIG. Of the density change binarization circuit 92 for binarization, the binary image memory 94 for storing the result, and the R image and G image data in which the density unevenness described in (a) above is corrected. A histogram calculation circuit 74, a threshold value calculation circuit 76, a density binarization circuit 78, and a binary image memory 80 that use image data are included. The specific configuration of the density change binarization circuit 92 is the same as the configuration shown in any of FIGS. 25, 26, 27, and 28 described in detail in the second embodiment. The binarization circuit 96 divides the area by logical OR (logical OR) between the data stored in the binary image memory 80 and the binary image memory 94 and stores the result in the binary image memory 98. .
[0226]
Note that the processing including the histogram calculation circuit 74, the threshold value calculation circuit 76, and the density binarization circuit 78 is executed in parallel with the processing of the density change binarization circuit 92, and can be speeded up. .
[0227]
The case where the density of the G image and the R image data is used has been described above, but the present invention is not limited to this configuration, and the two image data are appropriately selected from the R, G, and B image data. Using the selected image data, the threshold value calculation circuit 76 calculates a plurality of threshold values, and the image data binarized by the density binarization circuit 78 is stored in the binary image memory 80.
[0228]
Further, in the above description, the configuration of the density change binarization circuit 92 that detects the magnitude of the density change between the pixels of the G image data and the binary image memory 94 that stores the detected results have been described. A configuration including the change binarization circuit 92 and the binary image memory 94 is further connected in parallel to one or two, and in addition to the G image data, one image data of B image data or R image data, Alternatively, the two image data of B image data and R image data are configured to detect in parallel the magnitude of the change in density between pixels, and a plurality of density change binarizations connected in parallel in this way are arranged. A binarization circuit 96 performs a logical operation, for example, a logical sum, between the image detected from the circuit 92 and stored in the plurality of binary image memories 94 and the image stored in the binary image memory 80. Run and get The binarized image may be stored in the binary image memory 98.
[0229]
According to such a configuration, a plurality of images are used from the R, G, and B image data, and in each image, the magnitude of the change in density between pixels is detected and binarized. In order to obtain binarized image data by using a single image in R, G, B image data and detecting the magnitude of density change between pixels. In comparison, the particle region where the particle exists can be identified more accurately.
[0230]
(C) The apparatus which performs Example 3 has the configuration shown in FIG. That is, it has a configuration in which the above configurations (a) and (b) are combined. That is, a density change binarization circuit 92 that detects the magnitude of density change between pixels in the G image data corrected from density unevenness read out from the image memory 90 and obtains binarized image data; A binary image memory 94 for storing binarized image data, and a histogram calculation circuit 74 for obtaining a histogram of a G image and an R image using R image data and G image data in which density unevenness is corrected. Then, density binarization that obtains binarized image data from the G image and R image data using the threshold values T1, T2, T3, and T4 to detect the background region based on (Equation 9) A circuit 78, a binary image memory 80 for storing binarized image data, and a logical OR operation between the data stored in the binary image memory 80 and the binary image memory 94 2 A value circuit 96; Liwa a (logical OR) the result of the operation, a binary image memory 98 for storing the result of the area division, and a. The specific configuration of the density change binarization circuit 92 is the same as the configuration shown in any of FIGS. 25, 26, 27, and 28 described in detail in the second embodiment.
[0231]
Note that the processing including the histogram calculation circuit 74, the threshold value calculation circuit 76, and the density binarization circuit 78 is executed in parallel with the processing of the density change binarization circuit 92, and can be speeded up. .
[0232]
The case where the density of the G image and the R image data is used has been described above, but the present invention is not limited to this configuration, and the two image data are appropriately selected from the R, G, and B image data. Using the selected image data, the threshold value calculation circuit 76 calculates a plurality of threshold values, and the image data binarized by the density binarization circuit 78 is stored in the binary image memory 80. In addition, a plurality of configurations including a histogram calculation circuit 74, a threshold value calculation circuit 76, a density binarization circuit 78, and a binary image memory 80 are connected in parallel to further increase the density of the G image and R image data. For example, by using the density of the G image and B image data, a binarized image is obtained in a plurality of binary image memories 80, an image stored in the binary image memory 80, and a plurality of binaries A logic operation, for example, a logical sum, is executed by the binarization circuit 96 between the images stored in the image memory 80 and the obtained binarized image is stored in the binary image memory 98. Good.
[0233]
Further, in the above description, the configuration of the density change binarization circuit 92 that detects the magnitude of the density change between the pixels of the G image data and the binary image memory 94 that stores the detected results have been described. Further, one or two components including a density change binarization circuit 92 and a binary image memory 94 are connected in parallel so that one image data of B image data or R image data can be stored in addition to G image data. Alternatively, the two image data of B image data and R image data are configured to detect in parallel the magnitude of the change in density between pixels, and a plurality of density change binary values connected in parallel in this way A logical operation, for example, a logical sum, between the images detected from the conversion circuit 92 and stored in the plurality of binary image memories 80 and the image stored in the one or more binary image memories 94. Binary circuit 96 Run, binarized image obtained may be stored in the binary image memory 98.
[0234]
According to such a configuration, a plurality of images are used from the R, G, and B image data, and in each image, the magnitude of the change in density between pixels is detected and binarized. In order to obtain binarized image data by using a single image in R, G, B image data and detecting the magnitude of density change between pixels. In comparison, it is possible to more accurately identify the particle region where the particle exists.
[0235]
(D) The configuration of the apparatus that executes Example 4 includes the apparatus configurations of (a) and (b) described above, and was obtained from the apparatus configuration portions described in (a) and (b). The binarization circuit 96 performs a logical operation (logical sum) on the result, obtains a binary image divided into regions, and stores it in the binary image memory 98.
[0236]
Using the binary image divided into regions obtained in the binary image memory 98 having the apparatus configuration of (a) to (d) described above, the feature amount is calculated with respect to the target region where the particles are present. (Not shown in FIG. 32) is calculated and detected, and urine sediment particles are identified.
[0237]
【The invention's effect】
According to the present invention, in particular, in the region segmentation of an image obtained with a urine sediment examination apparatus that requires detection of target particles having a variety of properties, the target particles are not sufficiently stained due to the staining of the sample. Even in the case of a sample containing a large number of such target particles at a high frequency, the concentration of the particles can be extracted. That is, even when particles that are sufficiently stained or difficult to be stained (particles that are not sufficiently stained or particles that are hardly stained) are mixed due to the staining of the sample, each particle image can be divided stably and more accurately. As a result, the feature quantity of the target region can be obtained more accurately, the target particle classification error can be prevented, and the accuracy of identifying various sediment components can be improved.
[Brief description of the drawings]
FIG. 1 is a view for explaining the principle of a particle image image region dividing method according to a second embodiment of the present invention;
FIG. 2 is a view for explaining the principle of a particle image image region dividing method according to a second embodiment of the present invention;
FIG. 3 is a diagram showing a configuration example of a urine sediment examination apparatus to which the region dividing method of the present invention is applied.
FIG. 4 is a diagram schematically showing region division for extracting a background portion in a color space according to the present invention.
FIG. 5 is a diagram for explaining a density histogram for obtaining a parameter used to determine a threshold in the present invention.
FIG. 6 is a diagram showing an example of region division results of an image of a stained cell with good staining of target particles and an image of a cell with little staining of target particles in Example 1 of the present invention.
FIG. 7 is a diagram showing an example of region segmentation results of an image of a stained cell with good staining of target particles and an image of a cell with little staining of target particles in Example 1 of the present invention.
FIG. 8 is a diagram showing examples of density values of an image including cells that are well stained and an image including cells that are hardly stained in Example 2 of the present invention.
FIGS. 9A and 9B are diagrams showing examples of density values of an image including cells with good staining and an image including cells with little staining in Example 2 of the present invention. FIGS.
FIG. 10 is a diagram showing an example of dispersion values of an image including cells with good staining and an image including cells with little staining in Example 2 of the present invention.
FIGS. 11A and 11B are diagrams illustrating examples of dispersion values of an image including cells with good staining and an image including cells with little staining in Example 2 of the present invention. FIGS.
FIGS. 12A and 12B are diagrams showing examples of density difference values of an image including cells with good staining and an image including cells with little staining in Example 2 of the present invention.
FIG. 13 is a diagram showing examples of density difference values of an image including cells with good staining and an image including cells with little staining in Example 2 of the present invention.
FIG. 14 shows images including well-stained cells and cells that are hardly stained, using the binarization method using local dispersion of concentration in Example 2 of the present invention. The figure which shows the example of a result of area | region division | segmentation of an image.
FIG. 15 shows images including well-stained cells and cells that are hardly stained, using the binarization method using local dispersion of concentration in Example 2 of the present invention. The figure which shows the example of a result of area | region division | segmentation of an image.
FIGS. 16A and 16B are regions of an image including cells that are well stained and an image including cells that are hardly stained using the method based on binarization using density difference values in Example 2 of the present invention. The figure which shows the example of a result of a division | segmentation.
FIGS. 17A and 17B are regions of an image including cells that are well stained and an image including cells that are hardly stained using the method based on binarization using density difference values in Example 2 of the present invention. The figure which shows the example of a result of a division | segmentation.
FIG. 18 is a diagram illustrating an example of a result of area division according to a change in mask size according to the second embodiment of the present invention.
FIG. 19 is a diagram illustrating an example of a result of discrete Fourier transform in a background area and a target area in one scanning line in Embodiment 2 of the present invention.
FIG. 20 is a diagram illustrating an example of a frequency amplitude characteristic (| H (Ω) |) of a filter according to the second embodiment of the present invention.
FIG. 21 is a diagram showing an example of a result of examining the influence of a mask size on the shape of a divided region using an image of stained red blood cells in Example 2 of the present invention.
FIG. 22 is a flowchart of a process for explaining a procedure for dividing a region of a particle image in Embodiment 2 of the present invention.
FIG. 23 is a diagram showing a configuration example of an apparatus for carrying out a second embodiment of the present invention.
24 is a diagram showing a configuration example of an area dividing unit in FIG.
FIG. 25 is a diagram illustrating a specific circuit configuration example of a density change calculation unit according to the second embodiment of the present invention.
FIG. 26 is a diagram illustrating a specific circuit configuration example of a density change calculation unit according to the second embodiment of the present invention.
FIG. 27 is a diagram illustrating a specific circuit configuration example of a density change calculation unit according to the second embodiment of the present invention.
FIG. 28 is a diagram illustrating a specific circuit configuration example of a density change calculation unit according to the second embodiment of the present invention.
FIG. 29 is a diagram schematically illustrating the synthesis of two binary images divided into regions by two different methods according to the fourth embodiment of the present invention.
FIG. 30 is a diagram showing an example of the result of region division in an actual sample in Example 3 of the present invention.
FIG. 31 is a diagram showing an example of the result of region division in an actual sample in Example 3 of the present invention.
FIG. 32 is a diagram showing a configuration example of an apparatus to which the particle image image region dividing method of the present invention is applied.
[Explanation of symbols]
22 ... Urine sample, 24 ... Staining solution, 26 ... Flow cell, 28 ... Personal computer, 30 ... Flash lamp (Xe lamp), 32 ... Semiconductor laser, 34 ... Particle detector, 36 ... TV camera, 40 ... Density of R image Histogram, 40 '... G image density histogram, 50, 51, 52 ... Cell, 60, 61, 62 ... Noise component, 70 ... Correction data memory, 72 ... Density correction circuit, 74 ... Histogram calculation circuit, 76 ... Threshold Calculation circuit 78 ... Density binarization circuit, 80 ... Binary image memory, 90 ... Image memory, 92 ... Density change binarization circuit, 94 ... Binary image memory, 96 ... Binary circuit, 98 ... Binary Image memory, 102 ... TV camera, 103 ... A / D converter, 101 ... Pulse lamp, 102 ... TV camera, 103 ... A / D converter, 104 ... Image memory, 10 ... threshold value calculation unit, 106 ... area division unit, 107 ... image memory, 108 ... computer, 109 ... CPU bus, 110 ... data bus, 200 ... input unit, 210, 220, 230 ... binary image, 211, 212, 221 , 222, 231, 232 ... particle image, 300 ... density binarization circuit, 310 ... threshold memory, 320 ... comparator, 400 ... density change binarization circuit, 410 ... memory, 411-414, 420 ... density change Calculation unit, 421, 422 ... adder, 423 ... subtractor, 430 ... comparison unit, 431 ... threshold memory, 432 ... comparator, 500 ... logical sum circuit, 600 ... filter unit, 700 ... output unit, 800 ... variance calculation Circuit, 811 to 819 ... Memory, 821 ... Square calculation unit, 822, 824 ... Adder, 823, 825 ... Divider, 826 ... Square calculator, 827 ... Subtractor, 90 ... weighted sum calculation circuit, 931 to 939 ... memory, 941 to 949 ... weight value memory, 951-959 ... multiplier, 961 ... adder, 1000 ... image memory.

Claims (5)

A red component image (R image), a green component image (G image), and a blue component image (B image) of a stained target particle obtained by an image input optical system that inputs particle images are stationary. An image of a particle image that is captured and stored as an image, and that uses one or more of the R image, the G image, and the B image to identify a background region and a target region where the target particle exists. A region dividing method, wherein the background region and the density value of each image of a first group of one or more images selected from the R image, the G image, and the B image; A first step of setting a first group threshold for dividing a region having a large density difference and obtaining a binarized first binary image from the first group image; and the R image , A second composed of one or a plurality selected from the G image and the B image The amount of change in density value calculated in each pixel is calculated for each pixel by calculating the amount representing the magnitude of the change in density value in the vicinity region of each pixel constituting each image. A threshold value of the second group for dividing the region having a large change in the density value compared to the change in the density value in the background region, to an amount representing the size of the background region, from the image of the second group, A second step of obtaining a binarized second binary image, a logical operation of the first binary image and the second binary image, and the background region and the target region And a third step of obtaining a binarized third binary image representing the region of the particle image. 2. The particle image region dividing method according to claim 1, wherein the particle image is an image of stained urinary sediment particles scattered in a flow. A red component image (R image), a green component image (G image), and a blue component image (B image) of a stained target particle obtained by an image input optical system that inputs particle images are stationary. An image of a particle image that is captured and stored as an image, and that uses one or more of the R image, the G image, and the B image to identify a background region and a target region where the target particle exists. A first step of correcting a density value unevenness due to distortion of the image input optical system included in a plurality of images among the R image, the G image, and the B image; A region having a large density difference from the background region is divided with respect to the density value of each of the first group of images selected from the image, the G image, and the B image. Setting a first group threshold for the first group A second step of obtaining a binarized first binary image from the image, and a second group of images consisting of one or more selected from the R image, the G image, and the B image The amount representing the magnitude of the change in the density value in the vicinity region of each pixel constituting each image is calculated for each pixel, and the magnitude of the change in the density value obtained for each pixel. A threshold value of the second group for dividing an area where the change of the density value is large compared to the change of the density value in the background area is set as an amount representing the brightness, and binarization is performed from the image of the second group OR operation of the third step, the first binary image, and the second binary image, performed in parallel with the second step, to obtain a second binary image obtained A third step of obtaining a binarized third binary image representing the background area and the target area by: Segmentation methods of particle images which is characterized by having a. An image input optical system for inputting a particle image and a red component image (R image), a green component image (G image), and a blue component image (B image) of the stained target particles are obtained as still images. Storage means for storing image data of the R image, the G image, and the B image, and one or a plurality of images among the R image, the G image, and the B image are stored. A particle image region dividing device for identifying a background region and a target region in which the target particles are present, one or more selected from the R image, the G image, and the B image For the density value of each image of the first group of images, a threshold value of the first group for dividing the background region and a region having a large density difference is set. First binarization means for obtaining a binarized first binary image; The change of the density value in the vicinity region of each pixel constituting each image of the second group of one or more images selected from the R image, the G image, and the B image The amount representing the magnitude of the density value is calculated for each pixel, and the amount representing the magnitude of the change in the density value obtained for each pixel is compared with the change in the density value in the background region. A second binarization unit that sets a second group threshold value for dividing a region having a large change, and obtains a binarized second binary image from the second group image; A third binarization for obtaining a binarized third binary image representing the background region and the target region by a logical operation of the binary image of 1 and the second binary image Means for segmenting a particle image. An image input optical system for inputting a particle image and a red component image (R image), a green component image (G image), and a blue component image (B image) of the stained target particles are obtained as still images. Storage means for storing image data of the R image, the G image, and the B image, and one or a plurality of images among the R image, the G image, and the B image are stored. A particle image region dividing device that uses and distinguishes a background region and a target region where the target particles exist, and is included in a plurality of images among the R image, the G image, and the B image A means for correcting density value unevenness due to distortion of the image input optical system; and one or more selected from the R image, the G image, and the B image that have been corrected for the density value unevenness Each of the images of the first group of images For the density value, a first group threshold value for dividing the background area and the area having a large density difference is set, and a binarized first binary image is obtained from the first group image. The first binarization means to be obtained, and the vicinity of each pixel constituting each image of the second group of one or more images selected from the R image, the G image, and the B image An amount representing the magnitude of the change in the density value in the region is calculated for each pixel, and the amount representing the magnitude of the change in the density value obtained in each pixel is converted into the density in the background region. Setting a second group threshold value for dividing a region having a large change in density value compared with a change in value, and obtaining a binarized second binary image from the second group image, Second binarizing means operating in parallel with the first binarizing means; the first binary image; and And a third binarization means for obtaining a binarized third binary image representing the background region and the target region by a logical sum operation with the binary image of Particle image segmentation device.

Images