JP2004174219A - Apparatus and method for processing image and recording medium for storing program used for causing computer to execute the method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image processing apparatus etc., that objectively, efficiently, and automatically extracts the outline of an object of interest in, for example, an image for medical use. <P>SOLUTION: The image processing apparatus performs an optimum outline extraction on an image, from which an area of interest, such as a diagnosing portion, etc., can be observed appropriately, by smoothing a rough initial outline by using two different self-systemizing maps of a near learning area using the X- and Y-coordinate information of the initial outline as learning data. <P>COPYRIGHT: (C)2004,JPO

Description

【００４８】
（４）ユークリッド距離ｄ_ｊが最小となる出力ユニットを検索する。
【００４９】
（５）Ｎ_ｃ（ｔ）で定義される近傍に含まれるユニットへの重み係数を、次の式（２）にて更新する。
ｗ_ｉｊ（ｔ＋１）＝ｗ_ｉｊ（ｔ）＋α（ｔ）（ｘ_ｉ（ｔ）−ｗ_ｉｊ（ｔ））（２）
ここで、α（ｔ）は学習率係数（０＜α＜１）、Ｎ_ｃ（ｔ）は近傍領域のサイズであり、時間とともに減少させる。
【００５０】
（６）上記（２）〜（５）の処理を繰り返す。
【００５１】
本実施形態においては、初期輪郭線と上矢状静脈洞のエッジ部分とを合成して生成される脳の大まかな輪郭線を構成する複数の点（画素）の一部又は全部を、当該患者の脳の輪郭線を代表する代表点と見なす。この代表点を学習データとするＳＯＭ処理を繰り返し実行することで、より客観性の高い代表点を取得する。特に、以下に説明する例は、二種類のＳＯＭ処理を実行しその結果を合成することで、対象の複雑さに応じた輪郭抽出を実現するものである。二種類のＳＯＭ処理の一方は「初期近傍領域が３／４のＳＯＭ処理」と呼ばれ、他方は「初期近傍領域が１／８のＳＯＭ処理」と呼ばれる。
【００５２】
初期近傍領域が３／４のＳＯＭ処理では、図３に示すように、学習データ上の所定の各画素に関するＸ座標及びＹ座標を入力層における入力とし、所定の各重み係数によって、マッピング層における各出力ユニットに写像し、その中から最適な出力ユニット（勝者ユニット）を選択するＳＯＭ処理が繰り返し実行される。その際、各ＳＯＭ処理においては、学習近傍領域内に存在する出力ユニットに対応する各重み係数の更新が実行される。ここで、学習近傍領域とは、得られた勝者ユニットを少なくとも含む領域であり、学習回数の進行に従って収束するものである。また、初期近傍領域とは、学習近傍領域の初期設定を意味する。本ＳＯＭ処理では、初期近傍領域を、出力ユニット総数の３／４として設定し、所定の直線又は曲線に従って収束させる。
【００５３】
定量的な例により、上記初期近傍領域が３／４のＳＯＭ処理の内容をより詳しく説明すると、次のようである。すなわち、学習データ上の所定画素の座標（ｘ_ｉ，ｙ_ｉ）を入力とし、当該所定画素に対して式（１）に基づいてユークリッド距離ｄ_ｊを算出する。ここで、ｉは１≦ｉ≦最大学習回数（１０００００回）を満たす整数、ｊは１≦ｊ≦マッピング層の出力ユニット総数（１０００ユニット）を満たす整数である。算出されたユークリッド距離ｄ_ｊが最小となるユニットを勝者ユニットとし、勝者ユニット及び当該勝者ユニットを中心とした近傍領域内の出力ユニット（近傍ユニット）に対して式（２）に基づいて重み係数を更新する。今の場合、初期近傍領域が３／４であるから、当該領域内に存在する出力ユニットの総数は７５０ユニットとなる。
【００５４】
このように学習近傍領域を収束させつつ、当該領域内に存在する出力ユニットに対応する重み係数を式（２）によって逐次更新しながら、学習データ上の各画素についてのＳＯＭ処理を複数回（例えば１０００００回）実行する。最終的に、学習データ上の各画素の座標は、それぞれ所定の出力ユニットに収束するようになる。
【００５５】
また、初期近傍領域が１／８のＳＯＭ処理では、図３に示すように、初期近傍領域を最初に出力ユニット総数の１／８として設定し、所定の直線又は曲線に従って収束させながら、同様の処理をおこなう。
【００５６】
なお、本実施形態では、図４（ａ）に示すように、各ＳＯＭ処理における複数の出力ユニットは、第１出力ユニット（最初の出力ユニット）と第１０００出力ユニット（最後の出力ユニット）とが連結された閉空間を構成している。これにより、第１０００出力ユニットは、第１出力ユニットの近傍ユニットとして判断され、式（２）による重み係数の更新に影響することになる。この様に閉空間の出力ユニットを採用するのは、抽出しようとする輪郭が閉曲線であるからである。これにより、第１出力ユニットと第１０００出力ユニットとの相関度合いが高くなり、より精度の高い結果を得ることができる。
【００５７】
また、処理の収束性の観点から、学習回数が進行するに従って、各ＳＯＭ処理における近傍領域のサイズ及び学習率係数を線形で減少させることが好ましい。
【００５８】
（組織カラーマップの作成）
本画像処理装置１０は、輪郭抽出の前段の処理として、組織カラーマップを作成し、組織カラーマップ記憶部１２内に格納する。この組織カラーマップは、上記ＳＯＭ処理を利用した組織の自動分類法により、例えば次のように生成される。
【００５９】
すなわち、図４（ｂ）に示すように、組織分類の対象となるＭＲ画像（以下、「分類対象画像」と称する。図７（ａ）参照。）に対して、ランダムに抽出される複数の局所ブロックに関する４つの特徴量（注目画素の輝度値、局所ブロックを構成する各画素の輝度値の平均値、注目画素の輝度値と局所ブロックを構成する画素の輝度値の最大値との差、注目画素の輝度値と局所ブロックを構成する画素の輝度値の最小値との差）を入力層における入力とし、所定の重み係数によって患者の所定部位の組織分布を分類対象画像の特徴空間における画像の位相特性として自己写像する。この自己写像により、入力としての各局所ブロックは、ＳＯＭ処理により背景（ＢＧ）、脳脊髄液（ＣＳＦ）、白質（ＷＭ）、灰白質（ＧＭ）の４つの出力ユニットＡ、Ｂ、Ｃ及びＤのいずれかに写像される。
【００６０】
すなわち、いまｘ１＝注目画素の輝度値、ｘ２＝局所ブロックを構成する各画素の輝度値の平均値、ｘ３＝注目画素の輝度値と局所ブロックを構成する画素の輝度値の最大値との差、ｘ４＝注目画素の輝度値と局所ブロックを構成する画素の輝度値の最小値との差、とする。各局所ブロックにおける各特徴量ｘ_ｉ（ただし、ｉは１≦ｉ≦４を満たす整数）を入力し、式（１）に基づいて、初期設定の重み係数ｗ_ｉｊによってユークリッド距離ｄ_ｊを算出する。なお、今の場合、ｊ＝１が出力ユニットＣＳＦに対応、ｊ＝２が出力ユニットＧＭに対応、ｊ＝３が出力ユニットＷＭ、ｊ＝４が出力ユニットＢＧに対応するものとして、ｊは１≦ｊ≦４を満たす整数とする。
【００６１】
入力は、算出されたユークリッド距離ｄ_１、ｄ_２、ｄ_３、ｄ_４のうち、最小のものに対応する出力ユニットに写像される。なお、当該自己写像においてユークリッド距離が最小になる出力ユニットは、勝者ユニットと呼ばれる。
【００６２】
また、この抽出された局所ブロックに対するＳＯＭ処理は、式（２）によって逐次重み係数を更新しながら、複数回（例えば１００００００回）実行される。この様に、重み係数を更新しながら、ＳＯＭ処理を複数回繰り返し学習する。学習が進むに従って、各局所ブロックは、所定の出力ユニットに収束するようになる。
【００６３】
このＳＯＭ処理の結果と、図７（ａ）に示す分類対象画像において生成される、各画素を注目画素とする複数の局所ブロック（３×３画素）の４つの特徴量とに基づいて、分類対象画像の各画素を背景（ＢＧ）、脳脊髄液（ＣＳＦ）、白質（ＷＭ）、灰白質（ＧＭ）のいずれかに組織分類し、各画素に所定の色を割り当てることで、図７（ｂ）に示す組織カラーマップを生成することができる。
【００６４】
次に、上記のように構成した画像処理装置１０によって実行される輪郭抽出処理について、図５、図６を参照しながら説明する。
【００６５】
図５、図６は、画像処理装置１０が実行する輪郭抽出処理の流れを示したフローチャートである。図５において、まず、前段の処理として、既述の手法により組織カラーマップを作成し、組織カラーマップ記憶部１２に格納する（ステップＳ１）。学習フェーズ１３の前処理部１３０は、組織カラーマップ記憶部１２から図７（ｂ）に示す組織カラーマップを読み出し、所定の閾値処理によって当該組織カラーマップを脳組織と背景とに２値化した２値化画像を生成する（図７（ｃ）参照）。また、前処理部１３０は、生成した２値化画像に対して、脳組織のＸＹ射影から脳の仮想中心Ｃ（Ｃ_ｘ，Ｃ_ｙ）を算出する（ステップＳ２）。
【００６６】
次に、前処理部１３０は、仮想中心Ｃから脳組織の上半球を対象にフィルター処理を施し（ステップＳ３）、仮想中心Ｃから画像上方向に走査して、輪郭始点Ｓを求め、８近傍オペレータにより脳の大まかな初期輪郭を抽出する（ステップＳ４）。ここで、８近傍オペレータによる初期輪郭の抽出とは、例えば、輪郭始点Ｓに接する８個の画素に関して、画素値（輝度値）が最も近い画素を輪郭と判断し、この輪郭と判断された画素に接する８個の画素に関しても、同様の判断を逐次繰り返して輪郭を抽出するものである。すなわち、輪郭始点Ｓを中心とする第1近傍（３×３）内の画素に対して、背景領域に隣接する脳組織上の画素を順次抽出する処理である。
【００６７】
次に、前処理部１３０は、上矢状静脈洞のエッジ情報を抽出（ステップＳ５）、当該エッジ情報と抽出した上記初期輪郭とを合成し、学習データを生成する（図３参照）。自己学習部１３１は、生成された学習データ上の各画素のＸ座標ｘ_ｉ及びＹ座標ｙ_ｉをＳＯＭに入力し、既述の初期近傍領域が３／４のＳＯＭ処理及び初期近傍領域が１／８のＳＯＭ処理を実行し、各ＳＯＭ処理における各勝者ユニットを決定する（ステップＳ７）。このＳＯＭ処理による自己学習は、式（２）に従って逐次重み係数を更新しながら、信頼できる結果が得られる適当な回数（例えば、１０００００回）繰り返される（ステップＳ８）。
【００６８】
次に、重みプロット部１７０は、上記初期近傍領域が３／４のＳＯＭ処理によって得られた各重み係数を画像空間上の対応する座標にプロットし、各座標間を線形補間部１７１が線形補間することで当該初期近傍領域３／４の輪郭線を生成する。また、同様の処理により、初期近傍領域１／８の輪郭線を生成する（ステップＳ９）。これにより、各ＳＯＭ処理によって得られた高精度の代表点に基づく各輪郭線が得られる。
【００６９】
次に、統合フェーズ１９の基準点抽出部１９０は、図８（ａ）に示す初期近傍領域３／４の輪郭線と図８（ｂ）に示す初期近傍領域１／８の輪郭線とを図８（ｃ）に示す様に重ね合わせ（ステップＳ１０）、初期近傍領域１／８の輪郭線上において、大脳縦裂特徴点を基準点Ｄとして抽出する（ステップＳ１１）。
【００７０】
次に、輪郭線統合部１９１は、基準点抽出部１９０によって抽出された基準点Ｄに基づいて、初期近傍領域１／８の輪郭線と初期近傍領域３／４の輪郭線とを合成し脳輪郭線を生成する（ステップＳ１２、ステップＳ１３）。すなわち、輪郭線統合部１９１は、図９に示す様に、上記基準点Ｄ（大脳縦裂特徴点）から輪郭線を左右に走査し、初期近傍領域３／４の輪郭線との一つ目の左交点をＬ_１、二つ目の左交点をＬ_２とし、初期近傍領域３／４の輪郭線との一つ目の右交点をＲ_１、二つ目の右交点をＲ_２とする。得られたＬ_２−Ｒ_２間の輪郭線は初期近傍領域１／８の輪郭線とし、それ以外は初期近傍領域３／４の輪郭線として合成することで、脳輪郭線を生成する。なお、この様にＬ_２−Ｒ_２間の領域を初期近傍領域１／８の輪郭線とするのは、複雑な輪郭を持つ大脳縦裂特徴点に対応する領域を、より微細な構造を反映することができる輪郭初期近傍領域１／８の輪郭線で表すためである。また、Ｌ_２−Ｒ_２間以外の領域を初期近傍領域３／４の輪郭線とするのは、比較的複雑な輪郭を持たない領域を、曲線のスムーズな初期近傍領域３／４の輪郭線で表すためである。
【００７１】
本実施形態に係る画像処理装置によれば、個々の画像特性から大まかな初期輪郭を抽出し学習データとすることにより、組織の境界があいまいな画像に対しても最適な輪郭を、迅速かつ簡便に自動抽出することができる。従って、医師等に対して有益な診断情報を効率的に提供することができ、診断の質の向上に寄与することができる。
【００７２】
また、本実施形態に係る画像処理装置では、ＳＯＭ処理によって得られる内部的な基準に基づいて、輪郭の抽出が行われる。すなわち、従来の様に輪郭抽出のための基準ベクトルを外部に持つ必要がないことから、対象画像毎に複雑なパラメータの主観的設定を必要としない。従って、客観性の高い輪郭抽出を自動的に実行することができる。
【００７３】
実施形態において初期近傍領域が３／４のＳＯＭ処理と初期近傍領域が１／８のＳＯＭ処理とを実行した様に、本画像処理装置は、学習近傍領域及びマッピング層ユニット数を任意に設定することができる。従って、輪郭線のスムージング度合いを任意に設定することができ、その結果組織の複雑さの程度に合わせた最適な輪郭抽出を実現することができる。
【００７４】
本実施形態に係る画像処理装置では、マッピング層ユニットは、組織の境界領域線を多角形近似する頂点となっている。従って、再帰的処理により組織の境界があいまいでかつ不連続な画像において連続する境界線を抽出することができる。
【００７５】
以上、本発明を実施形態に基づき説明したが、本発明の思想の範疇において、当業者であれば、各種の変更例及び修正例に想到し得るものであり、それら変形例及び修正例についても本発明の範囲に属するものと了解される。例えば以下に示すように、その要旨を変更しない範囲で種々変形可能である。
【００７６】
本実施形態で説明した輪郭抽出処理は、当該方法をコンピュータに実行させるプログラムによって、パーソナルコンピュータやワークステーション等によっても実現することができる。この場合、コンピュータに当該方法を実行させることのできるプログラムは、磁気ディスク（フロッピーディスク、ハードディスクなど）、光ディスク（ＣＤ−ＲＯＭ、ＤＶＤなど）、半導体メモリなどの記録媒体に格納して頒布することができる。
【００７７】
また、各実施形態は可能な限り適宜組み合わせて実施してもよく、その場合組合わせた効果が得られる。さらに、上記実施形態には種々の段階の発明が含まれており、開示される複数の構成要件における適宜な組合わせにより種々の発明が抽出され得る。例えば、実施形態に示される全構成要件から幾つかの構成要件が削除されても、発明が解決しようとする課題の欄で述べた課題が解決でき、発明の効果の欄で述べられている効果の少なくとも１つが得られる場合には、この構成要件が削除された構成が発明として抽出され得る。
【００７８】
【発明の効果】
以上本発明によれば、例えば医用画像において、客観的基準かつ効率的に、関心対象の輪郭を自動的に抽出する画像処理装置、画像処理方法、及び当該画像処理をコンピュータに実行させるプログラムを格納する記録媒体を実現できる。
【図面の簡単な説明】
【図１】図１は、本実施形態に係る画像処理装置１０のブロック構成図を示している。
【図２】図２は、一般的な自己組織化マップ（ＳＯＭ）処理を説明するための図である。
【図３】図３は、本実施形態に係る画像処理装置が実施する自己組織化マップ（ＳＯＭ）処理を説明するための図である。
【図４】図４（ａ）は、本実施形態に係る画像処理装置が実施する自己組織化マップ（ＳＯＭ）処理を説明するための図である。図４（ｂ）は、組織の自動分類法において実施される自己組織化マップ（ＳＯＭ）処理を説明するための図である。
【図５】図５は、画像処理装置１０が実行する輪郭抽出処理の流れを示したフローチャートである。
【図６】図６は、画像処理装置１０が実行する輪郭抽出処理の流れを示したフローチャートである。
【図７】図７（ａ）は、組織の自動分類法の対象とする分類対象画像（ＭＲ画像）を示した写真である。図７（ｂ）は、組織の自動分類法によって得られる組織カラーマップを示した写真である。図７（ｃ）は、組織カラーマップを２値化して得られる２値化画像を示した写真である。
【図８】図８（ａ）は、初期近傍領域３／４の輪郭線を模式的に示した図である。図８（ｂ）は、初期近傍領域１／８の輪郭線を模式的に示した図である。図８（ｃ）は、初期近傍領域３／４の輪郭線と初期近傍領域１／８の輪郭線とを重ね合わせた画像を模式的に示した図である。
【図９】図９は、図８（ｃ）の基準点Ｄの近傍の拡大図である。
【符号の説明】
１０…画像処理装置
１１…ＭＲＩデータベース
１２…組織カラーマップ記憶部
１３…学習フェーズ
１５…ＳＯＭ重み記憶部
１７…生成フェーズ
１９…統合フェーズ
２１…脳表面輪郭線画像記憶部
１３０…前処理部
１３１…自己学習部
１７０…重みプロット部
１７１…線形補間部
１９０…基準点抽出部
１９１…輪郭線統合部
１９２…画像合成部[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to an image processing device incorporated in a medical image diagnostic device such as a magnetic resonance imaging device, or provided by a medical workstation or the like.
[0002]
[Prior art]
Medical images provide a great deal of clinical information about a subject in the form of images, and play an important role in many medical activities such as disease diagnosis, treatment and surgical planning. The medical image is embedded in the medical imaging device in data collected by a medical imaging device (for example, an ultrasound diagnostic device, an X-ray CT device, a nuclear medicine diagnostic device, a magnetic resonance diagnostic device, or the like), or It is generated by performing predetermined image processing by an image processing device realized by a medical workstation or the like.
[0003]
In the diagnosis using medical images, in order to realize a more objective diagnosis, for example, an MR image is tissue-classified utilizing characteristics of organs and lesions, and evaluation for quantitative image diagnosis of the form is performed. Law needs to be established. Taking the extraction of the outline of brain tissue in the diagnosis of brain atrophy as an example, since the brain tissue has large individual differences and cannot be extracted with uniform processing, the contour line of interest is extracted manually. At present, it is based on the processing of. Representative methods conventionally used include a contour extraction method using a snake model and a contour extraction method using a neural edge detector. The contour extraction method using the snake model has the advantage that, when a grayscale image is given, even if the originally appearing edges are missing or cut off, the contours of the brain surface can be extracted while restoring them. . In addition, the contour extraction method using a neural edge detector is a method in which a layered neural network learns an ideal edge image and an input image created from a contour traced by a specialist such as a physician. There is an advantage that contour extraction that matches well can be realized.
[0004]
However, it is difficult to automate the extraction processing by the contour extraction method using the snake model. This is because the parameters that determine the balance between the restriction on the density of the image and the restriction on the contour shape of the extraction target require fine adjustment for each image. Also, it is difficult to create a statistical model of the initial contour, and the extraction accuracy greatly varies depending on the contents of the initial contour model given by the operator, and therefore, the objectivity is poor.
[0005]
Also, in the contour extraction method using a neural edge detector, the content of an ideal edge image (teacher data) given as teacher data differs for each specialist such as a doctor. Poor sex. In addition, since the contour extraction accuracy of the method depends on the feature amount input to the hierarchical neural network and the network configuration (the number of intermediate layer units), setting for each image to be extracted is required, and the operator is instructed. Burden is great.
[0006]
The problems involved in these methods are not limited to the case where the diagnostic site is the brain, but also apply to a diagnostic site for which an objective reference and an efficient contour extraction method have not been established.
[0007]
[Problems to be solved by the invention]
The present invention has been made in view of the above circumstances, and includes, for example, an image processing apparatus, an image processing method, and an image processing method for automatically extracting an outline of a target of interest in a medical image, objectively and efficiently. It is an object of the present invention to provide a recording medium for storing a program to be executed by a computer.
[0008]
[Means for Solving the Problems]
The present invention employs the following means to achieve the above object.
[0009]
According to a first aspect of the present invention, there is provided a storage unit configured to store a first image related to a predetermined part of a subject, and a first contour extracting a first contour line of the predetermined part based on the first image. A contour extracting means, and a plurality of output units and a winner unit determined from the plurality of output units by associating the position of each pixel constituting the first contour with an input based on an individual weighting coefficient; Self-mapping means for obtaining the weight coefficient corresponding to each of the output units present in the learning neighborhood area based on the winner unit, and repeatedly executing the self-mapping updating while converging the learning neighborhood area. Based on each of the weighting factors corresponding to the plurality of output units acquired by the self-mapping that is repeatedly executed. An image processing apparatus characterized by comprising a second image generating means for generating a second image including a contour line, the.
[0010]
According to a second aspect of the present invention, in the apparatus according to the first aspect, the self-mapping unit is configured to repeatedly execute the self-mapping using the initial setting of the learning neighborhood area as a first size. And a second self-mapping unit that repeatedly executes the self-mapping with the initial setting of the learning neighborhood area set to a second size, wherein the second image generation unit includes the first self-mapping unit. The second contour line based on the weighting factors corresponding to the plurality of output units obtained by the second self-mapping unit and the weighting factors corresponding to the plurality of output units obtained by the second self-mapping unit. And generating the second image including the following.
[0011]
According to a third aspect of the present invention, in the device according to the first or second aspect, the number of the plurality of output units can be arbitrarily set.
[0012]
According to a fourth aspect of the present invention, in the device according to the first or second aspect, the size of the learning neighborhood area can be arbitrarily set.
[0013]
A fifth aspect of the present invention is the device according to any one of the first to fourth aspects, wherein the plurality of output units form a closed space.
[0014]
According to a sixth aspect of the present invention, in the device according to any one of the first to fifth aspects, the learning neighborhood area converges linearly as the self-mapping progresses.
[0015]
A seventh viewpoint of the present invention is a first contour line extracting step of extracting a first contour line of the predetermined part based on a first image related to a predetermined part of the subject, and the first contour line. The input of the position of each pixel that constitutes, and obtains a plurality of output units and a winner unit determined from the plurality of output units by association based on individual weight coefficients, wherein the winner unit The self-mapping step of repeatedly executing a self-mapping that updates the weighting coefficients corresponding to each of the output units present in the learning neighborhood area based on the self-mapping while converging the learning neighborhood area, and the self-mapping repeatedly executed. A second image generating a second image including a second contour line of the predetermined part based on each of the obtained weighting factors corresponding to the plurality of output units. An image processing method characterized by comprising the image generation step.
[0016]
According to an eighth aspect of the present invention, in the method according to the seventh aspect, the self-mapping step is a first self-mapping step of repeatedly executing the self-mapping with an initial setting of the learning neighborhood area being a first size. And a second self-mapping step of repeatedly executing the self-mapping by setting the initial setting of the learning neighborhood area to a second size. In the second image generation step, the first self-mapping is performed. The second contour based on the weighting factors corresponding to the plurality of output units obtained by the means and the weighting factors corresponding to the plurality of output units obtained by the second self-mapping means. Generating the second image including a line.
[0017]
According to a ninth aspect of the present invention, in the method according to the seventh or eighth aspect, the number of the plurality of output units can be arbitrarily set.
[0018]
According to a tenth aspect of the present invention, in the method according to the seventh or eighth aspect, the size of the learning neighborhood area can be arbitrarily set.
[0019]
An eleventh aspect of the present invention is the method according to any one of the seventh to tenth aspects, wherein the plurality of output units form a closed space.
[0020]
According to a twelfth aspect of the present invention, in the method according to any one of the seventh to eleventh aspects, the learning neighborhood area linearly converges as the self-mapping progresses.
[0021]
According to a thirteenth aspect of the present invention, a first contour line extracting step of extracting a first contour line of the predetermined part based on a first image related to the predetermined part of the subject, The input of the position of each pixel constituting the contour line of, to obtain a plurality of output units and a winner unit determined from the plurality of output units by association based on an individual weight coefficient, A self-mapping step of repeatedly executing a self-mapping that updates the weighting coefficients corresponding to each of the output units present in the learning neighborhood area based on the winner unit while converging the learning neighborhood area; and A second contour including a second contour line of the predetermined portion based on each of the weighting factors corresponding to the plurality of output units acquired by self-mapping; A computer-readable recording medium recording an image processing program for executing a second image generation step of generating an image.
[0022]
A fourteenth aspect of the present invention is the recording medium according to the thirteenth aspect, wherein the self-mapping step is a first self-mapping that repeatedly executes the self-mapping with the initial setting of the learning neighborhood area as a first size. And a second self-mapping step of repeatedly executing the self-mapping with the initial setting of the learning neighborhood area set to a second size. In the second image generation step, the first self-mapping step is performed. The second weighting coefficient corresponding to the plurality of output units obtained by the mapping means and the weighting coefficient corresponding to the plurality of output units obtained by the second self-mapping means; Generating the second image including an outline.
[0023]
According to a fifteenth aspect of the present invention, in the recording medium according to the thirteenth or fourteenth aspect, the number of the plurality of output units can be arbitrarily set.
[0024]
According to a sixteenth aspect of the present invention, in the recording medium according to the thirteenth or fourteenth aspect, the size of the learning neighborhood area can be arbitrarily set.
[0025]
A seventeenth aspect of the present invention is the recording medium according to any one of the thirteenth to sixteenth aspects, wherein the plurality of output units form a closed space.
[0026]
According to an eighteenth aspect of the present invention, in the recording medium according to any one of the thirteenth to seventeenth aspects, the learning neighborhood area converges linearly as the self-mapping progresses. .
[0027]
According to such a configuration, for example, in a medical image, an image processing apparatus and an image processing method for automatically extracting an outline of a target of interest in an objective standard and efficiently, and a program for causing a computer to execute the image processing are provided. A storage medium for storing can be realized.
[0028]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, components having substantially the same functions and configurations are denoted by the same reference numerals, and repeated description will be made only when necessary.
[0029]
The image processing apparatus according to the present embodiment targets an image in which a region of interest, such as a diagnostic region, can be suitably observed, and uses different X and Y coordinate information of rough initial contours as learning data. The rough initial contour is smoothed by using one Self-Organizing Map (hereinafter referred to as "SOM") to extract an optimal contour. In the following, in order to make the description specific, the T ₂ In the emphasized horizontal section, the rough initial contour and the X and Y coordinates of the upper sagittal sinus edge information are used as learning data, and the initial contour in which the longitudinal cerebral fissure feature point and the boundary of the tissue are ambiguous is smoothed. The case where the surface contour is automatically extracted is taken as an example.
[0030]
However, the technical idea of the present invention is not limited to this. For example, the image to be subjected to contour extraction may be another tomographic image such as an X-ray CT image, and it is needless to say that the image is useful in diagnosis other than the brain.
[0031]
FIG. 1 is a block diagram of an image processing apparatus 10 according to the present embodiment. As shown in FIG. 1, the image processing apparatus 10 includes an MRI database 11, a tissue color map storage unit 12, a learning phase 13, an SOM weight storage unit 15, a generation phase 17, an integration phase 19, and a brain surface contour image storage unit 21. Is provided.
[0032]
The MRI database 11 stores a plurality of MR images for each subject (patient).
[0033]
The tissue color map storage unit 12 stores a tissue color map obtained by applying an automatic tissue classification method to an MR image to be used for diagnosis. The automatic tissue classification method is to classify images for each tissue by self-mapping image characteristics of MR images to be used for diagnosis. The automatic classification method of this organization will be described later in detail.
[0034]
The learning phase 13 includes a pre-processing unit 130 and a self-learning unit 131. The preprocessing unit 130 performs predetermined preprocessing to set learning data as an input of the SOM. Specifically, the preprocessing unit 130 generates a binarized image in which the tissue color map is classified into a brain tissue and a background, calculates a virtual center C of the brain, and targets the upper hemisphere of the brain tissue from the virtual center C. Filter processing, extraction of an initial contour by an 8-neighbor operator, extraction of edge information of the superior sagittal sinus, and generation of learning data for SOM processing. The learning data is input data of the SOM, and is a rough outline of the brain generated by synthesizing the edge information with the initial outline extracted by the 8-neighbor operator (see FIG. 3).
[0035]
The self-learning unit 131 repeats self-learning with a predetermined weighting coefficient based on the learning data generated by the preprocessing unit 130, and learns the coordinates of the tissue contour and the coordinates of the edge information of the superior sagittal sinus. That is, the self-learning unit 131 receives the X and Y coordinates on the learning data as input, sequentially updates the weighting factor of the phase structure, sets the initial neighborhood area to the 1/8 neighborhood area, and executes the SOM processing. SOM processing is performed for a region near 3/4. The mapping layer to which each SOM is mapped is a closed space in which the first unit and the last unit are connected as neighbors. The SOM executed by the self-learning unit 131 will be described later in detail.
[0036]
The SOM weight storage unit 15 stores the weight coefficient distribution of the phase structure obtained by the SOM processing.
[0037]
The generation phase 17 includes a weight plot unit 170 and a linear interpolation unit 171. The weight plotting unit 170 converts the weighting factor obtained by the SOM processing of the initial neighborhood area by と and the weighting factor obtained by the SOM processing of the initial neighborhood area by ／ into image scales, respectively. Plot the coordinates. The linear interpolation unit 171 linearly interpolates the representative points plotted on the X and Y coordinates and generates an outline corresponding to each SOM processing (hereinafter, an outline corresponding to the ８ neighborhood area SOM processing is referred to as “initial The outline corresponding to the 1/4 neighborhood area SOM processing and the outline corresponding to the 3/4 neighborhood area SOM process are referred to as the "outline of the initial neighborhood area 3/4."
[0038]
The integration phase 19 includes a reference point extraction unit 190, a contour line integration unit 191, and an image synthesis unit 192. The reference point extraction unit 190 superimposes the contour line of the initial neighborhood region 1/8 and the contour line of the initial neighborhood region 3/4, and extracts, for example, a longitudinal cerebral fissure feature point on the contour line of the initial neighborhood region 1/8. I do. Based on the reference point extracted by the reference point extraction unit 190, the contour line of the initial neighborhood area 1/8, and the contour line of the initial neighborhood area 3/4, the contour line integration unit 191 generates the contour of the initial neighborhood area 1/8. The line and the contour of the initial neighborhood 3/4 are integrated to generate a brain contour image.
[0039]
The image synthesizing unit 192 generates a synthesized image obtained by synthesizing an MR image (original image of a tissue color map) and a brain contour image as needed.
[0040]
The brain surface outline image storage unit 21 stores the synthesized image, the brain outline image, and the like generated by the image synthesis unit 192, and outputs them to a display unit (not shown) or the like as required.
[0041]
In FIG. 1, the MRI database 11, the tissue color map storage unit 12, and the SOM weight storage unit 15 are individually represented. However, a configuration in which each configuration is realized by a single memory may be used. Further, the MRI database 11 and the tissue color map storage unit 12 may be provided separately from the image processing apparatus 10 to acquire an MR image and a tissue color map via a network.
[0042]
(Self-Organizing Map)
Next, a self-organizing map (SOM) will be described. SOM preserves the phase of input data and performs topological mapping, and does not require an explicit teacher in the learning process.
[0043]
As shown in FIG. 2, a general SOM has two layers: an input layer including an input unit, and a mapping layer including an output unit. A typical SOM learning algorithm is as follows.
[0044]
(1) w _ij Let (1 ≦ i ≦ n) be the weighting factor from input unit i to output unit j at time t. The weight coefficient of the unit is initialized with random numbers, and the initial range near node j is set large.
[0045]
(2) x _i (1 ≦ i ≦ n) is an input to the node i at the time t.
[0046]
(3) Euclidean distance d between input data and output node j _j Is calculated by the following equation (1).
[0047]
(Equation 1)

[0048]
(4) Euclidean distance d _j Find the output unit that minimizes.
[0049]
(5) N _c The weight coefficient for the unit included in the neighborhood defined by (t) is updated by the following equation (2).
w _ij (T + 1) = w _ij (T) + α (t) (x _i (T) -w _ij (T)) (2)
Here, α (t) is a learning rate coefficient (0 <α <1), N _c (T) is the size of the neighborhood area, which decreases with time.
[0050]
(6) The above processes (2) to (5) are repeated.
[0051]
In the present embodiment, a part or all of a plurality of points (pixels) constituting a rough outline of the brain generated by synthesizing the initial outline and the edge portion of the superior sagittal sinus are determined by the patient. Is considered as a representative point representing the outline of the brain. By repeatedly executing the SOM process using this representative point as learning data, a more objective representative point is obtained. In particular, the example described below implements contour extraction according to the complexity of the target by executing two types of SOM processing and synthesizing the results. One of the two types of SOM processing is referred to as “SOM processing in which the initial neighborhood area is ／”, and the other is referred to as “SOM processing in which the initial neighborhood area is ８”.
[0052]
In the SOM processing in which the initial neighborhood area is 3/4, as shown in FIG. 3, the X-coordinate and the Y-coordinate of each predetermined pixel on the learning data are input to the input layer, and the predetermined weighting coefficients are used for the mapping layer. SOM processing for mapping to each output unit and selecting an optimum output unit (winner unit) from the output units is repeatedly executed. At that time, in each SOM process, updating of each weight coefficient corresponding to an output unit existing in the learning neighborhood area is executed. Here, the learning neighborhood area is an area including at least the obtained winner unit, and converges as the number of times of learning progresses. Further, the initial neighborhood area means the initial setting of the learning neighborhood area. In the present SOM processing, the initial neighborhood area is set as ／ of the total number of output units, and is converged according to a predetermined straight line or curve.
[0053]
The details of the SOM processing in which the initial neighborhood area is 3/4 will be described in more detail with reference to a quantitative example. That is, the coordinates (x _i , Y _i ) Is input, and the Euclidean distance d is calculated for the predetermined pixel based on Expression (1). _j Is calculated. Here, i is an integer satisfying 1 ≦ i ≦ maximum number of learning times (100,000 times), and j is an integer satisfying 1 ≦ j ≦ the total number of output units of the mapping layer (1000 units). The calculated Euclidean distance d _j Is defined as the winner unit, and the weight coefficient is updated based on the equation (2) for the winner unit and the output unit (neighboring unit) in the vicinity area centered on the winner unit. In this case, since the initial neighborhood area is 3/4, the total number of output units existing in the area is 750 units.
[0054]
As described above, the SOM processing for each pixel on the learning data is performed a plurality of times (for example, while the weighting coefficient corresponding to the output unit existing in the learning area is converged while the weighting coefficient corresponding to the output unit existing in the area is sequentially updated by the equation (2)). 100000 times). Eventually, the coordinates of each pixel on the learning data converge on a predetermined output unit.
[0055]
Also, in the SOM processing in which the initial neighborhood area is 1/8, as shown in FIG. 3, the initial neighborhood area is first set as 1/8 of the total number of output units, and converged according to a predetermined straight line or curve. Perform processing.
[0056]
In the present embodiment, as shown in FIG. 4A, a plurality of output units in each SOM process include a first output unit (first output unit) and a 1000th output unit (last output unit). It forms a connected closed space. As a result, the 1000th output unit is determined as a neighboring unit of the first output unit, and affects the updating of the weight coefficient according to equation (2). The reason why the output unit of the closed space is adopted in this way is that the contour to be extracted is a closed curve. Thereby, the degree of correlation between the first output unit and the 1000th output unit is increased, and a more accurate result can be obtained.
[0057]
Further, from the viewpoint of the convergence of the process, it is preferable to linearly decrease the size of the neighboring region and the learning rate coefficient in each SOM process as the number of times of learning progresses.
[0058]
(Creating an organization color map)
The image processing apparatus 10 creates a tissue color map and stores it in the tissue color map storage unit 12 as a process prior to the outline extraction. This tissue color map is generated, for example, as follows by an automatic tissue classification method using the above-described SOM processing.
[0059]
That is, as shown in FIG. 4B, a plurality of randomly extracted MR images (hereinafter, referred to as “classification target images”; see FIG. 7A) are targeted for tissue classification. Four feature amounts relating to the local block (the luminance value of the pixel of interest, the average value of the luminance values of the pixels forming the local block, the difference between the luminance value of the pixel of interest and the maximum value of the luminance values of the pixels forming the local block, The difference between the luminance value of the target pixel and the minimum luminance value of the pixels constituting the local block) is used as an input in the input layer, and the tissue distribution of a predetermined part of the patient in a feature space of the classification target image is determined using a predetermined weighting coefficient. Self-mapping as the phase characteristic of By this self-mapping, each local block as input is converted into four output units A, B, C and D of the background (BG), cerebrospinal fluid (CSF), white matter (WM) and gray matter (GM) by SOM processing. Is mapped to any of
[0060]
That is, x1 = the luminance value of the target pixel, x2 = the average value of the luminance values of the pixels forming the local block, x3 = the difference between the luminance value of the target pixel and the maximum luminance value of the pixels forming the local block. , X4 = the difference between the luminance value of the pixel of interest and the minimum luminance value of the pixels forming the local block. Each feature x in each local block _i (Where i is an integer that satisfies 1 ≦ i ≦ 4), and based on the equation (1), the initially set weight coefficient w _ij By the Euclidean distance d _j Is calculated. In this case, j = 1 corresponds to the output unit CSF, j = 2 corresponds to the output unit GM, j = 3 corresponds to the output unit WM, j = 4 corresponds to the output unit BG, and j is 1 It is an integer satisfying ≦ j ≦ 4.
[0061]
The input is the calculated Euclidean distance d ₁ , D ₂ , D ₃ , D ₄ Are mapped to the output unit corresponding to the smallest one. Note that the output unit that minimizes the Euclidean distance in the self-mapping is called a winner unit.
[0062]
The SOM processing on the extracted local block is executed a plurality of times (for example, 1,000,000 times) while sequentially updating the weighting coefficient by the equation (2). As described above, the SOM processing is repeatedly performed a plurality of times while updating the weight coefficient. As the learning progresses, each local block converges on a predetermined output unit.
[0063]
Based on the result of the SOM processing and four feature amounts of a plurality of local blocks (3 × 3 pixels) each of which is a pixel of interest generated in the classification target image shown in FIG. Tissue classification of each pixel of the target image into one of the background (BG), cerebrospinal fluid (CSF), white matter (WM), and gray matter (GM), and assigning a predetermined color to each pixel, FIG. The tissue color map shown in b) can be generated.
[0064]
Next, the contour extraction processing executed by the image processing apparatus 10 configured as described above will be described with reference to FIGS.
[0065]
FIG. 5 and FIG. 6 are flowcharts illustrating the flow of the contour extraction processing executed by the image processing apparatus 10. In FIG. 5, first, as a preceding process, a tissue color map is created by the above-described method and stored in the tissue color map storage unit 12 (step S1). The preprocessing unit 130 of the learning phase 13 reads the tissue color map shown in FIG. 7B from the tissue color map storage unit 12, and binarizes the tissue color map into a brain tissue and a background by a predetermined threshold process. A binarized image is generated (see FIG. 7C). Further, the preprocessing unit 130 applies a virtual center C (C) to the generated binarized image from the XY projection of the brain tissue. _x , C _y ) Is calculated (step S2).
[0066]
Next, the preprocessing unit 130 performs a filtering process on the upper hemisphere of the brain tissue from the virtual center C (step S3), scans the image upward from the virtual center C, finds the contour start point S, A rough initial contour of the brain is extracted by the operator (step S4). Here, extraction of the initial contour by the 8-neighbor operator refers to, for example, regarding eight pixels in contact with the contour start point S, a pixel having the closest pixel value (luminance value) is determined to be a contour, and a pixel determined to be the contour is determined. The same determination is repeated for the eight pixels in contact with. That is, for pixels in the first neighborhood (3 × 3) centered on the contour start point S, pixels on the brain tissue adjacent to the background region are sequentially extracted.
[0067]
Next, the preprocessing unit 130 extracts edge information of the superior sagittal sinus (step S5), and combines the edge information with the extracted initial contour to generate learning data (see FIG. 3). The self-learning unit 131 calculates the X coordinate x of each pixel on the generated learning data. _i And Y coordinate y _i Is input to the SOM, and the SOM processing in which the above-described initial neighborhood area is 3/4 and the SOM processing in which the initial neighborhood area is 1/8 are executed, and each winner unit in each SOM processing is determined (step S7). The self-learning by the SOM process is repeated an appropriate number of times (for example, 100,000 times) to obtain a reliable result while sequentially updating the weighting coefficient according to the equation (2) (step S8).
[0068]
Next, the weight plotting unit 170 plots the respective weighting coefficients obtained by the SOM processing in which the initial neighborhood area is 3/4 on the corresponding coordinates in the image space, and the linear interpolation unit 171 linearly interpolates between the coordinates. By doing so, a contour line of the initial neighborhood area 3/4 is generated. Further, a contour line of the initial neighborhood area 1/8 is generated by the same processing (step S9). Thereby, each contour line based on the highly accurate representative points obtained by each SOM process is obtained.
[0069]
Next, the reference point extraction unit 190 of the integration phase 19 draws a contour line of the initial neighboring area 3/4 shown in FIG. 8A and a contour line of the initial neighboring area 1/8 shown in FIG. As shown in FIG. 8 (c), the cerebral longitudinal fissure feature point is extracted as a reference point D on the contour line of the initial neighborhood area 1/8 (step S11).
[0070]
Next, based on the reference point D extracted by the reference point extraction unit 190, the contour line integration unit 191 combines the contour of the initial neighborhood area 1/8 and the contour of the initial neighborhood area 3/4, and An outline is generated (step S12, step S13). That is, as shown in FIG. 9, the contour line integration unit 191 scans the contour line from the reference point D (the cerebral longitudinal fissure feature point) to the left and right, and compares the contour line with the contour line of the initial neighboring area 3/4. L at the left intersection of ₁ , The second left intersection is L ₂ And the first right intersection with the contour line of the initial neighborhood area 3/4 is R ₁ , The second right intersection is R ₂ And L obtained ₂ -R ₂ The outline between them is the outline of the initial neighborhood area 1/8, and the rest is synthesized as the outline of the initial neighborhood area 3/4, thereby generating the brain outline. Note that L ₂ -R ₂ The region between them is defined as the contour line of the initial neighborhood region 1/8. The region corresponding to the longitudinal cerebral fissure feature point having a complicated contour is defined as the contour initial neighborhood region 1 / that can reflect a finer structure. This is because it is represented by eight contour lines. Also, L ₂ -R ₂ The reason for defining the area other than the area as the contour of the initial neighborhood area 3/4 is to represent the area having no relatively complicated contour by the contour of the smooth initial neighborhood area 3/4.
[0071]
According to the image processing apparatus according to the present embodiment, a rough initial contour is extracted from individual image characteristics and used as learning data, so that an optimum contour can be quickly and easily obtained even for an image having an ambiguous tissue boundary. Can be automatically extracted. Therefore, useful diagnostic information can be efficiently provided to doctors and the like, which can contribute to improvement in the quality of diagnosis.
[0072]
Further, in the image processing apparatus according to the present embodiment, contour extraction is performed based on an internal reference obtained by the SOM processing. That is, since there is no need to have a reference vector for contour extraction outside as in the related art, there is no need to subjectively set complicated parameters for each target image. Therefore, highly objective contour extraction can be automatically executed.
[0073]
As in the embodiment, the image processing apparatus arbitrarily sets the number of learning neighboring regions and the number of mapping layer units, as in the case where the SOM processing in which the initial neighboring region is ／ and the SOM processing in which the initial neighboring region is １ ／ are performed in the embodiment. be able to. Therefore, it is possible to arbitrarily set the degree of smoothing of the contour, and as a result, it is possible to realize optimal contour extraction according to the degree of complexity of the tissue.
[0074]
In the image processing device according to the present embodiment, the mapping layer unit is a vertex that approximates the boundary area line of the tissue with a polygon. Therefore, a continuous boundary line can be extracted in a discontinuous image in which the boundary between tissues is ambiguous and discontinuous by the recursive processing.
[0075]
As described above, the present invention has been described based on the embodiments. However, in the scope of the concept of the present invention, those skilled in the art can come up with various modified examples and modified examples. It is understood that it belongs to the scope of the present invention. For example, as shown below, various modifications can be made without changing the gist.
[0076]
The contour extraction processing described in the present embodiment can also be realized by a personal computer, a workstation, or the like by a program that causes a computer to execute the method. In this case, the program that can cause the computer to execute the method can be stored in a recording medium such as a magnetic disk (floppy disk, hard disk, and the like), an optical disk (CD-ROM, DVD, and the like), a semiconductor memory, and distributed. it can.
[0077]
In addition, the embodiments may be implemented in appropriate combinations as much as possible, in which case the combined effects can be obtained. Further, the embodiments include inventions at various stages, and various inventions can be extracted by appropriately combining a plurality of disclosed constituent elements. For example, even if some components are deleted from all the components shown in the embodiment, the problem described in the column of the problem to be solved by the invention can be solved, and the effects described in the column of the effect of the invention can be solved. When at least one of the above is obtained, a configuration from which this configuration requirement is deleted can be extracted as an invention.
[0078]
【The invention's effect】
According to the present invention, for example, an image processing apparatus and an image processing method for automatically extracting an outline of a target of interest in an objective standard and efficiently in a medical image, and a program for causing a computer to execute the image processing are stored. Recording medium that can be realized.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating an image processing apparatus 10 according to an embodiment.
FIG. 2 is a diagram for explaining a general self-organizing map (SOM) process;
FIG. 3 is a diagram for explaining a self-organizing map (SOM) process performed by the image processing apparatus according to the embodiment;
FIG. 4A is a diagram illustrating a self-organizing map (SOM) process performed by the image processing apparatus according to the embodiment. FIG. 4B is a diagram for describing a self-organizing map (SOM) process performed in the automatic tissue classification method.
FIG. 5 is a flowchart illustrating a flow of an outline extraction process performed by the image processing apparatus 10;
FIG. 6 is a flowchart illustrating a flow of an outline extraction process performed by the image processing apparatus 10;
FIG. 7A is a photograph showing a classification target image (MR image) to be subjected to the automatic tissue classification method. FIG. 7B is a photograph showing a tissue color map obtained by an automatic tissue classification method. FIG. 7C is a photograph showing a binarized image obtained by binarizing the tissue color map.
FIG. 8A is a diagram schematically illustrating an outline of an initial vicinity region 3/4. FIG. 8B is a diagram schematically illustrating the contour of the initial vicinity area 1/8. FIG. 8C is a diagram schematically showing an image in which the outline of the initial neighborhood area 3/4 and the outline of the initial neighborhood area 1/8 are superimposed.
FIG. 9 is an enlarged view near a reference point D in FIG. 8C.
[Explanation of symbols]
10 ... Image processing device
11 ... MRI database
12 ... Tissue color map storage
13 ... Learning phase
15 ... SOM weight storage unit
17… Generation phase
19 ... Integration phase
21: Brain surface contour image storage
130 ... Pre-processing unit
131 ... Self-learning unit
170 ... weight plot section
171: Linear interpolation unit
190: Reference point extraction unit
191 ... Contour line integration unit
192 ... Image synthesis unit

Claims (18)

Storage means for storing a first image relating to a predetermined part of the subject;
First contour line extraction means for extracting a first contour line of the predetermined part based on the first image;
Self-mapping in which the position of each pixel constituting the first contour line is input, and a plurality of output units and a winner unit determined from the plurality of output units are obtained by association based on individual weighting factors. Means, the weighting coefficients corresponding to each of the output units present in the learning neighborhood area based on the winner unit, self-mapping means repeatedly executing self-mapping updating while converging the learning neighborhood area,
A second image that generates a second image including a second contour line of the predetermined portion based on each of the weighting factors corresponding to the plurality of output units acquired by the self-mapping that is repeatedly executed. Generating means;
An image processing apparatus comprising: The self-mapping means includes first self-mapping means for repeatedly executing the self-mapping with the initial setting of the learning neighborhood area as a first size, and the self-mapping means with the initial setting of the learning neighborhood area as a second size. And a second self-mapping means for repeatedly executing
The second image generating means includes: a weighting coefficient corresponding to the plurality of output units obtained by the first self-mapping means; and a plurality of output units obtained by the second self-mapping means. Generating the second image including the second contour based on the corresponding weighting factor;
The image processing apparatus according to claim 1, wherein: The image processing apparatus according to claim 1, wherein the number of the plurality of output units can be arbitrarily set. 3. The image processing apparatus according to claim 1, wherein the size of the learning neighborhood area can be arbitrarily set. The image processing apparatus according to any one of claims 1 to 4, wherein the plurality of output units form a closed space. The image processing apparatus according to claim 1, wherein the learning neighborhood area converges linearly with the progress of the self-mapping. A first contour line extraction step of extracting a first contour line of the predetermined part based on a first image related to the predetermined part of the subject;
The position of each pixel constituting the first contour is input, and a plurality of output units and a winner unit determined from the plurality of output units are acquired by association based on individual weighting factors. Self-mapping step of repeatedly executing a self-mapping that updates the weighting coefficients corresponding to each of the output units present in the learning neighborhood area based on the winner unit while converging the learning neighborhood area,
A second image that generates a second image including a second contour line of the predetermined portion based on each of the weighting factors corresponding to the plurality of output units acquired by the self-mapping that is repeatedly executed. Generating step;
An image processing method comprising: The self-mapping step includes a first self-mapping step of repeatedly executing the self-mapping with the initial setting of the learning neighborhood area as a first size, and the self-mapping step with the initial setting of the learning neighborhood area as a second size. A second self-mapping step of repeatedly executing
In the second image generation step, the weight coefficients corresponding to the plurality of output units obtained by the first self-mapping unit and the plurality of output units obtained by the second self-mapping unit Generating the second image including the second contour line based on the weight coefficient corresponding to
The image processing method according to claim 7, wherein: 9. The image processing method according to claim 7, wherein the number of the plurality of output units can be arbitrarily set. 9. The image processing method according to claim 7, wherein the size of the learning neighborhood area can be arbitrarily set. The image processing method according to claim 7, wherein the plurality of output units form a closed space. The image processing method according to claim 7, wherein the learning neighborhood area linearly converges as the self-mapping progresses. On the computer,
A first contour line extraction step of extracting a first contour line of the predetermined part based on a first image related to the predetermined part of the subject;
The position of each pixel constituting the first contour is input, and a plurality of output units and a winner unit determined from the plurality of output units are acquired by association based on individual weighting factors. Self-mapping step of repeatedly executing a self-mapping that updates the weighting coefficients corresponding to each of the output units present in the learning neighborhood area based on the winner unit while converging the learning neighborhood area,
A second image that generates a second image including a second contour line of the predetermined portion based on each of the weighting factors corresponding to the plurality of output units acquired by the self-mapping that is repeatedly executed. Generating step;
And a computer-readable recording medium storing an image processing program for executing the program. The self-mapping step includes a first self-mapping step of repeatedly executing the self-mapping with the initial setting of the learning neighborhood area as a first size, and the self-mapping step with the initial setting of the learning neighborhood area as a second size. A second self-mapping step of repeatedly executing
In the second image generation step, the weight coefficients corresponding to the plurality of output units obtained by the first self-mapping unit and the plurality of output units obtained by the second self-mapping unit Generating the second image including the second contour line based on the weight coefficient corresponding to
14. The computer readable recording medium according to claim 13, wherein: 15. The computer-readable recording medium according to claim 13, wherein the number of the plurality of output units can be arbitrarily set. The computer-readable recording medium according to claim 13, wherein the size of the learning neighborhood area can be arbitrarily set. 17. The computer-readable recording medium according to claim 13, wherein the plurality of output units form a closed space. 18. The computer-readable recording medium according to claim 13, wherein the learning neighborhood area converges linearly with the progress of the self-mapping.

Images