JP2005020338A - Method, apparatus and program for detecting abnormal shadow - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method, apparatus and program for enhancing the detection performance of an abnormal shadow on the basis of a medical image. <P>SOLUTION: A structure normal image generating means 20 artificially generates a structure normal image Sh representing a normal structure of a prescribed structural object corresponding to a structural image representing a prescribed structural object configuring an object in a received medical image P, a structure-removed image generating means 30 subtracts the structure normal image Sh from the medical image P to generate a structure-removed image Rh from which the structure image causing a hindrance to the detection of the abnormal shadow is removed, and an abnormal shadow detection means 40 detects the abnormal shadow on the basis of the image Rh. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

ここで、ｘ，ｙは各サンプル画像中のランドマーク各点の座標、ｘ′，ｙ′はワーピングされる平均形状上の座標、Δｘ，Δｙは平均形状へのシフト量、ｎは次数、ａｉｊ，ｂｉｊは係数である。なお、多項式近似の係数は最小自乗法を用いて求める。
【００４７】
上式を用いて、各サンプル画像について画素毎の平均形状へのシフト量を算出し、各サンプル画像を画素毎に平均形状へワーピングする。このとき、ワーピング後の座標が整数ではなく小数点以下を含む位置に移動する画素については、４近傍から１次近似で画素値を求める。つまり、ワーピング後の座標を囲む４つの画素に対して、ワーピング後の座標から各画素の座標までの距離に比例して画素値をそれぞれ分配するようにする。そして、平均形状へワーピングされた後の各サンプル画像から平均テクスチャを作成する。
【００４８】
この時点で、胸部についての、平均形状、その平均形状へワーピングされた後の各サンプル画像、および平均テクスチャが得られる。
【００４９】
（２）形状変化モデルの作成（ステップＳ２）
各サンプル画像における胸部の形状とその平均形状を用いて、胸部の形状に対して主成分分析を行い、固有形状を算出する。固有成分は文献 ［Ｍａｔｔｈｅｗ Ｔｕｒｋ， Ａｌｅｘ Ｐｅｎｔｌａｎｄ， Ｅｉｇｅｎｆａｃｅｓ ｆｏｒ Ｒｅｃｏｇｎｉｔｉｏｎ， Ｊｏｕｒｎａｌ ｏｆ Ｃｏｇｎｉｔｉｖｅ Ｎｅｕｒｏｓｃｉｅｎｃｅ， ｖｏｌ．３， ｎｕｍ１， １９９１］に示す手法を用いて求める。固有形状の線形和を用いて任意形状が次式のように近似できる。
【００５０】
【数２】

ここで、ｘは形状ベクトル（ｘ１ ，ｙ１ ，・・・，ｘｉ ，ｙｉ ，・・・，ｘｎ ，ｙｎ ）、ｘａｖｅ は平均形状ベクトル、Ｐｓ は形状の固有ベクトル、ｂｓ は形状係数のセット（組み合わせ）である。図５は、固有形状の例として、互いに異なる固有形状Ｋ１，Ｋ２，Ｋ３をランドマークＭで表したものである、なお、図中、実線のＫａは平均形状である。
【００５１】
（３）テクスチャ変化モデルの作成（ステップＳ３）
各サンプル画像における胸部のテクスチャとその平均テクスチャを用いて、胸部のテクスチャに対して主成分分析を行い、固有テクスチャを算出する。
【００５２】
【数３】

ここで、ｇはグレイレベル（テクスチャの画素値）のベクトル（ｇ１１，ｇ２１，・・・，ｇｉｊ，・・・，ｇｎｎ）、ｇａｖｅ は平均グレイレベルのベクトル、Ｐｇ はグレイレベルの固有ベクトル、ｂｇ はグレイレベル係数のセットである。図６は、固有テクスチャの例として、互いに異なる固有テクスチャＴ１，Ｔ２，Ｔ３を示したものである。
【００５３】
（４）形状とテクスチャの相関モデルの作成（ステップＳ４）
各サンプル画像における胸部の形状／テクスチャは、各固有形状／固有テクスチャに投影した値からなる係数ｂｓ 、ｂｇ を用いて表現することができる。以下に示すように、これら係数に対してさらに主成分分析を行い、形状とグレイレベルをコントロールするアピアランスパラメータを算出する。
【００５４】
【数４】

ここで、Ｗｓ は対角行列で各形状係数に対する重みである。
【００５５】
【数５】

Ｑは固有ベクトル、ｃは形状とグレイレベルを制御するアピアランスパラメータ（ｃの平均は０）である。このアピアランスパラメータを変化させることにより、胸部の略任意の画像を人工的に生成することができる。
【００５６】
（５）正常画像の生成（ステップＳ５）
胸部画像Ｐに対応する胸部正常画像を、以下のような手続きで生成する。
１．平均形状と平均形状内の画像から算出したアピアランスパラメータを初期値として設定する（正常画像データ取得手段）。
２．上位の固有成分のアピアランスパラメータから順番にパラメータを少しずつ振りながら人工画像を作成し、その人工画像と胸部画像との差分が所定値以下となるパラメータを探索する（構造変更ベクトル設定手段，構造変更ベクトル加算手段）。
３．上記２．で求めたアピアランスパラメータの値から得られた人工画像を、胸部画像Ｐに対応する胸部正常画像として決定する。
【００５７】
上記手法によれば、予め、被写体の平均形状／平均テクスチャ、固有形状／固有テクスチャ、固有ベクトルＱといった各要素を求めておけば、アピアランスパラメータｃの調整のみによって、医用画像に対応する略任意の画像を生成することができる。
【００５８】
なお、被写体を構成する所定の構造物に対応する画像を生成するには、サンプル画像として、その所定の構造物のみを表す画像を用意すればよく、例えば、実際に撮影された被写体全体を表す画像から、その所定の構造物のみを表す領域だけを抽出した画像を生成し、サンプル画像として用いることができる。この領域の抽出は、テンプレートマッチングによる手法や濃淡の変化を調べてエッジを検出する手法などの画像認識技術を用いたり、あるいは、手動で切り抜いたりすることで実現できる。
【００５９】
（実施例１）
図７は、本発明の第１の異常陰影検出装置の一実施形態である第１異常陰影検出装置の構成を示した概略ブロック図である。図７に示す第１異常陰影検出装置は、解剖学的特徴を有する、肋骨や鎖骨等の骨部とそれ以外の軟部とからなる人体胸部を被写体とした胸部放射線画像Ｐ（以下、単に胸部画像という）を入力する被写体画像入力手段１０と、胸部画像Ｐにおける骨部を表す骨部画像Ｐｈに対応する、骨部の正常な構造を表す骨部正常画像Ｓｈを推測して人工的に生成する構造物正常画像生成手段２０と、胸部画像Ｐと骨部正常画像Ｓｈとの画像間演算により、胸部から骨部が除去された像を表す骨部除去画像Ｒｈを生成する構造物除去画像生成手段３０と、生成された骨部除去画像Ｒｈに基づいて異常陰影検出を行う異常陰影検出手段４０と、生成された骨部除去画像Ｒｈを異常陰影検出結果とともに画面に表示する表示手段５０とを備えている。
【００６０】
次に、上記のように構成された第１異常陰影検出装置の作用について説明する。図８は、当該装置の処理フローを示した図である。
【００６１】
被写体画像入力手段１０は、放射線画像をデジタル画像として取得可能なＣＲ装置や、放射線画像のデジタル画像が保存されている記憶装置等から、人体胸部の放射線画像を表す胸部画像Ｐを入力する（ステップＳ１１）。ここでは、胸部画像Ｐは、骨部の後ろにある癌化部分を表す異常陰影ｆが含まれているものとする。図９は、このような胸部画像Ｐを表した図である。
【００６２】
構造物正常画像生成手段２０は、入力された胸部画像Ｐに基づいて、下記の処理により、骨部正常画像Ｓｈを得る。
【００６３】
まず、胸部画像Ｐにおいて画像の濃淡の急峻な変化を検出する等により、骨のエッジすなわち骨部とそれ以外の軟部組織との境界を検出し、胸部画像Ｐを、骨部を表す骨部画像Ｐｈと軟部を表す軟部画像Ｐｇとに分割して認識する（ステップＳ１２）。
【００６４】
次に、認識された骨部画像Ｐｈに対応する、骨部の正常な構造を表す骨部正常画像Ｓｈと、認識された軟部画像Ｐｇに対応する、軟部の正常な構造を表す軟部正常画像Ｓｇとを暫定的に生成し（ステップＳ１３）、さらに、骨部正常画像Ｓｈと軟部正常画像Ｓｇとを合成して、胸部全体を表す統合画像Ｓｈｇを生成する（ステップＳ１４）。
【００６５】
骨部正常画像Ｓｈおよび軟部正常画像Ｓｇは、前述のＡＡＭの手法を用いて、正常な骨部のみを表すサンプル画像および正常な軟部のみを表すサンプル画像を教師データとして、骨部および軟部の正常画像を人工的に生成するための演算式をそれぞれ求め、各演算式に対して適当なアピアランスパラメータを探索することにより生成する。
【００６６】
そして、生成された統合画像Ｓｈｇと元の胸部画像Ｐとを比較して、画像のパターンが互いに充分に近いか否かを判定し（ステップＳ１５）、充分近い場合は、このときの骨部正常画像Ｓｈを骨部画像Ｐｈに対応する画像として確定し（ステップＳ１６）、そうでない場合には、骨部正常画像Ｓｈと軟部正常画像Ｓｇのいずれか一方あるいは両方を、上記アピアランスパラメータを変更することにより生成し直し（ステップＳ１３）、同様の処理を繰り返すことにより、骨部画像Ｐｈに対応した骨部正常画像Ｓｈを得る。なお、画像パターンが充分に近いか否かの判定は、例えば、比較対象となる２つの画像間で対応する画素同士の差分信号値の分散が所定値以内であるときに、充分近いと判定することができる。図１０は、このようにして得られた骨部正常画像Ｓｈを表した図である。
【００６７】
構造物除去画像生成手段３０は、胸部画像Ｐと構造物正常画像生成手段２０により得られた骨部正常画像Ｓｈとの間で画像間演算としての減算処理、具体的には、胸部画像Ｐから骨部正常画像Ｓｈを、対応する画素同士で引き算することにより、胸部から骨部が除去された像を表す骨部除去画像Ｒｈを生成する（ステップＳ１７）。図１１は、骨部除去画像Ｒｈを表した図である。このようにして得られた骨部除去画像Ｒｈには、胸部画像Ｐがもともと有していた軟部の情報と異常陰影の情報が残るので、エネルギーサブトラクション画像に相当する画像として利用することができる。
【００６８】
異常陰影検出手段４０は、生成された骨部除去画像Ｒｈに基づいて、既存の各種の異常陰影検出手法により、異常陰影を検出する（ステップＳ１８）。異常陰影検出手法としては、例えば、特開２００２−１０９５１０号公報にて提案されているアイリスフィルタを用いて腫瘤陰影を検出する手法などが挙げられる。
【００６９】
表示手段５０は、異常陰影検出結果とともに骨部除去画像Ｒｈを、ＣＲＴや液晶パネル等の画面に表示する（ステップＳ１９）。なお、表示手段５０は、骨部正常画像Ｓｈと元の胸部画像Ｐとを同一画面上で並べて表示したり、これら２つの画像を位置合せし重ね合せて表示したりしてもよい。
【００７０】
なお、実施例１においては、骨部正常画像Ｓｈおよび軟部正常画像Ｓｇを暫定的に生成する際に、認識された骨部画像Ｐｈおよび軟部画像Ｐｇと比較しながら画像パターンの近いものだけを採択し、それらを用いて統合画像Ｓｈｇを生成するようにしているが、認識された骨部画像Ｐｈおよび軟部画像Ｓｇとの比較は行わずに、各画像を生成する演算式において、アピアランスパラメータを適当に変更してゆき、その都度、統合画像Ｓｈｇを生成するようにしてもよい。この場合、骨部画像Ｐｈに対応する骨部正常画像Ｓｈを見つけて確定するまでの処理においては、骨部正常画像Ｓｈ、軟部正常画像Ｓｇ、統合画像Ｓｈｇは、必ずしも画像データとして生成する必要はなく、各画像に対するグレイレベル係数のセットｂｇ を、これらの画像を特定する情報として用いればよい。
【００７１】
また、骨部正常画像Ｓｈと軟部正常画像Ｓｇとを暫定的に生成し、これらを合成して統合画像Ｓｈｇを生成する処理においては、骨部正常画像Ｓｈを生成するための演算と、軟部正常画像Ｓｇを生成するための演算とは、実質的に別の演算になっていればどのような処理をしてもよく、これらを別々に処理してもよいし、見かけ上一体に処理してもよい。例えば、軟部のグレイレベルｕと骨部のグレイレベルｈとを、それぞれ上記の式（６）を利用して、
【数６】

と表したとすると、統合画像のグレイレベルを、
【数７】

として、見かけ上一体の処理として扱うようにしてもよい。
【００７２】
また、上記の実施例１においては、胸部画像Ｐから骨部正常画像Ｓｈを減算することにより、胸部から骨部が除去された像を表す骨部除去画像Ｒｈを得ているが、胸部画像Ｐから統合画像Ｓｈｇを減算することにより、胸部から骨部および軟部が除去された像を表す骨部軟部除去画像Ｒｈｇを得るようにしてもよい。このようにして得られた骨部軟部除去画像Ｒｈｇには、胸部画像Ｐがもともと有していた異常陰影の情報が残るので、経時サブトラクション画像に相当する画像として利用することができる。
【００７３】
（実施例２）
本発明の第１の異常陰影検出装置の他の実施形態である第２異常陰影検出装置について説明する。第２異常陰影検出装置は、上記第１異常陰影検出装置と同じ構成であり、図７に示すように、被写体画像入力手段１０と、構造物正常画像生成手段２０と、構造物除去画像生成手段３０と、異常陰影検出手段４０と、表示手段５０とを備えているが、構造物正常画像生成手段２０における正常画像の生成プロセスが異なる。
【００７４】
次に、第２異常陰影検出装置の作用について説明する。図１２は、当該装置の処理フローを示した図である。
【００７５】
被写体画像入力手段１０は、ＣＲ装置や記憶装置等から、人体胸部の放射線画像を表す胸部画像Ｐを入力する（ステップＳ２１）。
【００７６】
構造物正常画像生成手段２０は、入力された胸部画像Ｐに基づいて、下記の処理により、骨部正常画像Ｓｈを得る。
【００７７】
まず、胸部画像Ｐにおいて画像の濃淡の急峻な変化を検出する等により、骨のエッジすなわち骨部とそれ以外の軟部組織との境界を検出し、軟部を表す軟部画像Ｐｇを認識する（ステップＳ２２）。
【００７８】
次に、認識された軟部画像Ｐｇに対応する、軟部の正常な構造を表す軟部正常画像Ｓｇを生成し（ステップＳ２３）、胸部画像Ｐから軟部正常画像Ｓｇを、対応する画素同士で引き算することにより、胸部から軟部が除去された像を表す軟部除去画像Ｒｇを生成する（ステップＳ２４）。この時点で、軟部除去画像Ｒｇには、骨部の情報と異常陰影の情報が含まれている。さらに、軟部除去画像Ｒｇにおいて骨のエッジを検出し、軟部除去画像Ｒｇにおいて骨部を表す骨部画像Ｐｈ′を認識する（ステップＳ２５）。認識された骨部画像Ｐｈ′に対応する、骨部の正常な構造を表す骨部正常画像Ｓｈ′を生成する（ステップＳ２６）。
【００７９】
軟部正常画像Ｓｇおよび骨部正常画像Ｓｈ′は、前述のＡＡＭの手法を用いて、正常な軟部のみを表すサンプル画像および正常な骨部のみを表すサンプル画像を教師データとして、軟部および骨部の正常画像を人工的に生成するための演算式をそれぞれ求め、各演算式に対して適当なアピアランスパラメータを探索することにより生成する。
【００８０】
構造物除去画像生成手段３０は、構造物正常画像生成手段２０により得られた軟部除去画像Ｒｇと骨部正常画像Ｓｈ′との間で画像間演算としての減算処理、具体的には、軟部除去画像Ｒｇから骨部正常画像Ｓｈ′を、対応する画素同士で引き算することにより、胸部から軟部および骨部が除去された像を表す軟部骨部除去画像Ｒｇｈを生成する（ステップＳ２７）。このようにして得られた軟部骨部除去画像Ｒｇｈには、胸部画像Ｐがもともと有していた異常陰影の情報が残るので、経時サブトラクション画像に相当する画像として利用することができる。
【００８１】
異常陰影検出手段４０は、生成された軟部骨部除去画像Ｒｇｈに基づいて、上述の各種の異常陰影検出手法により、異常陰影を検出する（ステップＳ２８）。
【００８２】
表示手段５０は、異常陰影検出結果とともに軟部骨部除去画像Ｒｇｈを、ＣＲＴや液晶パネル等の画面に表示する（ステップＳ２９）。なお、表示手段５０は、骨部正常画像Ｓｈ′と元の胸部画像Ｐ、軟部正常画像Ｓｇと元の胸部画像Ｐ、あるいは、軟部骨部除去画像Ｒｇｈと元の胸部画像Ｐを、同一画面上で並べて表示したり、これら２つの画像を位置合せし重ね合せて表示したりしてもよい。
【００８３】
（実施例３）
本発明の第１の異常陰影検出装置のさらに別の実施形態である第３異常陰影検出装置について説明する。第３異常陰影検出装置は、上記第１異常陰影検出装置と同じ構成であり、図７に示すように、被写体画像入力手段１０と、構造物正常画像生成手段２０と、構造物除去画像生成手段３０と、異常陰影検出手段４０と、表示手段５０とを備えているが、構造物正常画像生成手段２０における正常画像の生成プロセスが異なる。
【００８４】
次に、第３異常陰影検出装置の作用について説明する。図１３は、当該装置の処理フローを示した図である。
【００８５】
被写体画像入力手段１０は、ＣＲ装置や記憶装置等から、人体胸部の放射線画像を表す胸部画像Ｐを入力する（ステップＳ３１）。
【００８６】
構造物正常画像生成手段２０は、入力された胸部画像Ｐに基づいて、下記の処理により、骨部正常画像Ｓｈを得る。
【００８７】
まず、胸部画像Ｐに対応する、胸部の正常な構造を表す胸部正常画像Ｓｋを生成する（ステップＳ３２）。胸部正常画像Ｓｋは、前述のＡＡＭによる手法を用いて、正常な胸部を表すサンプル画像を教師データとして、胸部の正常画像を人工的に生成するための演算式を求め、この演算式に対して適当なアピアランスパラメータを探索することにより生成する。
【００８８】
そして、胸部画像と骨部画像との対応関係を予め学習させておいたニューラルネットを用いて、胸部正常画像Ｓｋと対応関係にある骨部画像を求め、これを胸部画像Ｐにおける骨部画像Ｐｈに対応する、骨部の正常な構造を表す骨部正常画像Ｓｈとする（ステップＳ３３）。ここで、「胸部画像と骨部画像との対応関係」とは、例えば、胸部のグレイレベルｋと骨部のグレイレベルｈとを、それぞれ上記の式（６）を利用して、
【数８】

と表したとすると、ｂｋ とｂｈ の対応関係とすることができる。なお、ニューラルネットの詳細に関しては、特開２０００−４８１８７号公報に記載されているので参照されたい。
【００８９】
構造物除去画像生成手段３０は、胸部画像Ｐと構造物正常画像生成手段２０により得られた骨部正常画像Ｓｈとの間で画像間演算としての減算処理、具体的には、胸部画像Ｐから骨部正常画像Ｓｈを、対応する画素同士で引き算することにより、胸部から骨部が除去された像を表す骨部除去画像Ｒｈを生成する（ステップＳ３４）。このようにして得られた骨部除去画像Ｒｈには、胸部画像Ｐがもともと有していた軟部の情報と異常陰影の情報が残るので、エネルギーサブトラクション画像に相当する画像として利用することができる。
【００９０】
異常陰影検出手段４０は、生成された骨部除去画像Ｒｈに基づいて、上述の各種の異常陰影検出手法により、異常陰影を検出する（ステップＳ３５）。
【００９１】
表示手段５０は、異常陰影検出結果とともに骨部除去画像Ｒｈを、ＣＲＴや液晶パネル等の画面に表示する（ステップＳ３６）。
【００９２】
このような、本発明の第１の異常陰影検出装置の実施形態である第１から第３異常陰影検出装置によれば、被写体を構成する構造物の正常な構造を表す構造物正常画像を人工的に生成し、医用画像と構造物正常画像との画像間演算により、医用画像から構造物画像が除去された画像を得、当該画像に基づいて異常陰影を検出するので、従来のような単純なフィルタリング処理では充分に除去できなかった、異常陰影の検出に対して障害となる複雑な構造を有する背景特徴画像としての構造物画像を、医用画像の画像パターンに応じて抽出して除去することが可能となり、異常陰影の検出性能を向上させることができる。
【００９３】
なお、実施例１から３においては、所定の画像に対応する正常な構造を表す正常画像を得るのに、画像を人工的に生成する手法を用いているが、その他の手法として、予め人工的に生成された多数の画像の中から、画像パターンが略合致するものを選択する手法（本発明の第２の異常陰影検出方法）も考えられる。この手法を用いた場合の実施例について以下に説明する。
【００９４】
（実施例４）
図１４は、本発明の第２の異常陰影検出装置の一実施形態である第４異常陰影検出装置の構成を示した概略ブロック図である。図１４に示した第４異常陰影検出装置は、胸部画像Ｐを入力する被写体画像入力手段１０と、人工的に生成された、胸部における軟部および骨部の正常な構造を表す複数の異なる画像を保存する画像保存手段６０と、胸部画像Ｐにおける骨部画像Ｐｈに略合致する画像Ｃｈを、画像保存手段６０により保存されている画像の中から選択する画像選択手段７０と、胸部画像Ｐと選択された画像Ｃｈとの画像間演算により、胸部から骨部が除去された像を表す骨部除去画像Ｒｈを生成する構造物除去画像生成手段３０と、生成された骨部除去画像Ｒｈに基づいて異常陰影検出を行う異常陰影検出手段４０と、生成された骨部除去画像Ｒｈを異常陰影検出結果とともに画面に表示する表示手段５０とを備えている。また、この第４異常陰影検出装置は、実施例１における第１異常陰影検出装置に対応するものである。
【００９５】
次に、上記のように構成された第４異常陰影検出装置の作用について説明する。図１５は、当該装置の処理フローを示した図である。
【００９６】
被写体画像入力手段１０は、放射線画像をデジタル画像として取得可能なＣＲ装置や、放射線画像のデジタル画像が保存されている記憶装置等から、人体胸部の放射線画像を表す胸部画像Ｐを入力する（ステップＳ４１）。
【００９７】
画像選択手段７０は、入力された胸部画像Ｐに基づいて、下記の処理により、骨部正常画像Ｓｈを得る。
【００９８】
まず、胸部画像Ｐにおいて画像の濃淡の急峻な変化を検出する等により、骨のエッジすなわち骨部とそれ以外の軟部組織との境界を検出し、胸部画像Ｐを、骨部を表す骨部画像Ｐｈと軟部を表す軟部画像Ｐｇとに分割して認識する（ステップＳ４２）。
【００９９】
次に、認識された骨部画像Ｐｈに略合致する、骨部の正常な構造を表す骨部正常画像Ｓｈと、認識された軟部画像Ｐｇに略合致する、軟部の正常な構造を表す軟部正常画像Ｓｇとを、画像保存手段６０により保存されている骨部および軟部の多数の人工画像の中から暫定的に選択し（ステップＳ４３）、さらに、骨部正常画像Ｓｈと軟部正常画像Ｓｇとを合成して、胸部全体を表す統合画像Ｓｈｇを生成する（ステップＳ４４）。
【０１００】
そして、生成された統合画像Ｓｈｇと元の胸部画像Ｐとを比較して、画像のパターンが互いに充分に近いか否かを判定し（ステップＳ４５）、充分近い場合は、そのときの骨部正常画像Ｓｈを骨部画像Ｐｈに対応する画像として確定し（ステップＳ４６）、そうでない場合には、骨部正常画像Ｓｈと軟部正常画像Ｓｇのいずれか一方あるいは両方を選択し直し（ステップＳ４３）、同様の処理を繰り返すことにより、骨部画像Ｐｈに対応した骨部正常画像Ｓｈを得る。このとき、選択された人工画像を、そのまま骨部正常画像Ｓｈとして用いてもよいが、選択された人工画像を認識された骨部画像Ｓｈの形状にワーピングした画像を、骨部正常画像Ｓｈとして用いるようにしてもよい。なお、画像パターンが充分に近いか否かの判定は、例えば、比較対象となる２つの画像間で対応する画素同士の差分信号値の分散が所定値以内であるときに、充分近いと判定することができる。これ以降は、実施例１と同様なので省略する。
【０１０１】
このような、本発明の第２の異常陰影検出装置の一実施形態である第４異常陰影検出装置によれば、被写体を構成する構造物の正常な構造を表す画像に相当する画像を、人工的に生成された複数の画像の中から選択し、医用画像と選択された画像との画像間演算により、医用画像から構造物画像が除去された画像を得、当該画像に基づいて異常陰影を検出するので、本発明の第１の異常陰影検出方法を適用したシステムと同様に、従来のような単純なフィルタリング処理では充分に除去できなかった、異常陰影の検出に対して障害となる複雑な構造を有する背景特徴画像としての構造物画像を、医用画像の画像パターンに応じて抽出して除去することが可能となり、異常陰影の検出性能を向上させることができる。
【０１０２】
なお、上記実施例４は、実施例１における第１異常陰影検出装置に対応するものであるが、同様に、実施例２、３における第２、第３異常陰影検出装置に対応するものも考えることができる。
【０１０３】
上記の実施例１から４においては、最終的に、骨部除去画像、軟部骨部除去画像を生成しているが、同様にして、軟部除去画像を生成することも、もちろん可能である。
【０１０４】
なお、このようにして得られた構造物除去画像は、異常陰影検出以外にも用途があり、例えば、主に軟部を表す骨部除去画像は、含気の具合を可視化して、肺機能の診断に利用することができ、また、主に骨部を表す軟部除去画像は、骨密度の定量化など骨の診断に利用することもできる。
【０１０５】
（実施例５）
本発明の異常陰影検出装置は、種々の装置からなるシステムにネットワーク等を介して接続して用いることもできる。
【０１０６】
図１６は、そのようなシステムの一例である異常陰影検出処理システムの概略を示した構成図である。図１６に示すシステムは、被写体画像を表す画像データを取得するＣＲ装置やＣＴ装置、ＭＲＩ装置などの入力モダリティ８１と、入力モダリティ８１で取得された画像データを含む種々のデータを記憶する記憶装置８２と、画像データが表す被写体画像から所望の構造物が除去された構造物除去画像を表す構造物除去画像データを生成するとともに、生成された構造物除去画像データに基づいて、構造物除去画像における異常陰影を検出する異常陰影検出装置８３と、被写体画像や構造物除去画像、異常陰影検出結果等の種々の情報を画面に出力する表示装置８４と、それらの情報をハードコピーとして出力するフィルムプリンタ８５とから構成される。
【０１０７】
次に、上記のように構成された異常陰影検出処理システムの動作について説明する。
【０１０８】
入力モダリティ８１あるいは記憶装置８２から異常陰影検出処理装置８３に画像データを入力する。すると、異常陰影検出処理装置８３において、入力された画像データに基づいて、被写体画像から所望の構造物が除去された構造物除去画像を表す構造物除去画像データを生成するとともに、生成された構造物除去画像データに基づいて、構造物除去画像における異常陰影を検出する。そして、異常陰影検出結果を、単独で、もしくは原画像データや構造物除去画像データと組にして、記憶装置８２や表示装置８４、フィルムプリンタ８５等に出力する。記憶装置８２に記憶・保存する際には、異常陰影検出結果を原画像データや構造物除去画像データと対応付けて保存する。このとき、構造物除去画像データの代わりに、構造物除去画像データの生成を再現できるような情報（生成時のパラメータ）を保存するようにしてもよい。
【図面の簡単な説明】
【図１】正常画像の生成手順を表すフローチャートを示す図
【図２】胸部サンプル画像に対して肺野領域の外縁にランドマークを付した状態を示す図
【図３】胸部の平均形状を表す画像Ｐａを示す図
【図４】胸部サンプル画像において肺野領域の外縁に付されたランドマーク各点が平均形状へワーピングされる状態を示す図
【図５】互いに異なる固有形状をランドマークで表す図
【図６】互いに異なる固有テクスチャを示す図
【図７】第１異常陰影検出装置の構成を示すブロック図
【図８】第１異常陰影検出装置の処理フローを示す図
【図９】入力された胸部画像Ｐを示す図
【図１０】骨部正常画像Ｓｈを示す図
【図１１】骨部除去画像Ｒｈを示す図
【図１２】第２異常陰影検出装置の処理フローを示す図
【図１３】第３異常陰影検出装置の処理フローを示す図
【図１４】第４異常陰影検出装置の構成を示すブロック図
【図１５】第４異常陰影検出装置の処理フローを示す図
【図１６】異常陰影検出処理システムの構成を示す図
【符号の説明】
１０ 被写体画像入力手段
２０ 構造物画像生成手段
３０ 構造物除去画像生成手段
４０ 異常陰影検出手段
５０ 表示手段
６０ 画像保存手段
７０ 画像選択手段
８１ 入力モダリティ
８２ 記憶装置
８３ 異常陰影検出装置
８４ 表示装置
８５ フィルムプリンタ
Ｃｈ 選択された骨部の画像
ｆ 異常陰影
Ｋ１ 固有形状
Ｋ２ 固有形状
Ｋ３ 固有形状
Ｋａ 平均形状
Ｍ ランドマーク
Ｐ 入力された胸部画像
Ｐ１，Ｐ２，Ｐ３ 胸部サンプル画像
Ｐａ 平均形状画像
Ｐｈ（Ｐｈ′） 骨部画像
Ｐｇ 軟部画像
Ｓｈ（Ｓｈ′） 骨部正常画像
Ｓｇ 軟部正常画像
Ｓｈｇ 統合画像
Ｓｋ 胸部正常画像
Ｒｈ 骨部除去画像
Ｒｇ 軟部除去画像
Ｒｇｈ 軟部骨部除去画像
Ｔ１，Ｔ２，Ｔ３ 固有テクスチャ[0001]
BACKGROUND OF THE INVENTION
The present invention relates to abnormal shadow detection processing, and more particularly to an abnormal shadow detection method and apparatus suitable for automatic detection of a tumor shadow in a digital chest X-ray image, and a program therefor.
[0002]
[Prior art]
Conventionally, in the medical field, abnormal shadow detection processing (CAD) for automatically detecting an abnormal shadow in an image based on a digital medical image using a computer has been performed. There is a chest CAD process for detecting a mass shadow in an image of the chest based on an X-ray image.
[0003]
A chest X-ray image includes a so-called “background feature image” composed of images representing structures having various anatomical features such as ribs and clavicles. This background feature image is used to detect abnormal shadows. This is an obstacle and causes a decrease in detection ability.
[0004]
Therefore, Patent Document 1 proposes a technique of removing such background feature images by filtering processing and performing chest CAD processing.
[0005]
[Patent Document 1]
JP-A-6-121792
[0006]
[Problems to be solved by the invention]
However, the anatomical structure of the chest is complex, and the chest CAD processing using the above filtering processing has a problem that the background feature image cannot be sufficiently removed and the detection performance of abnormal shadows is not improved. It was.
[0007]
In view of the above circumstances, an object of the present invention is to provide an abnormal shadow detection method and apparatus capable of further improving abnormal shadow detection performance, and a program therefor.
[0008]
[Means for Solving the Problems]
The first abnormal shadow detection method of the present invention is a structure image representing at least one predetermined structure among the plurality of structures in a medical image representing a subject composed of a plurality of structures having anatomical features. The structure normal structure image representing the normal structure of the predetermined structure corresponding to the step is artificially generated, and the structure from the medical image is calculated by the inter-image calculation between the medical image and the generated structure normal image. A method comprising generating a structure removal image from which an image has been removed, and detecting an abnormal shadow in the structure removal image based on the structure removal image.
[0009]
The first abnormal shadow detection apparatus of the present invention includes an image input means for inputting a medical image representing a subject composed of a plurality of structures having anatomical features, and among the plurality of structures in the medical image. A structure normal image generating means for artificially generating a structure normal image representing a normal structure of the predetermined structure corresponding to a structure image representing at least one predetermined structure, and a medical image are generated. The structure removal image generating means for generating the structure removal image in which the structure image is removed from the medical image by the inter-image calculation with the normal structure image, and the structure removal image based on the structure removal image An abnormal shadow detecting means for detecting an abnormal shadow is provided.
[0010]
In the first abnormal shadow detection method, the step of generating a structure normal image includes the step of acquiring image data corresponding to an image representing a normal structure of a structure of the same type as the predetermined structure, The step can include a step of setting a plurality of structure change vectors for changing the structure of the structure, and a step of adding the plurality of structure change vectors to the image data.
[0011]
Similarly, in the first abnormal shadow detection apparatus, the structure normal image generating means acquires image data corresponding to an image representing a normal structure of a structure of the same type as the predetermined structure. A structure change vector setting means for setting a plurality of structure change vectors for changing the structure of the same kind of structure, and a structure change vector addition means for adding the plurality of structure change vectors to the image data Can be provided.
[0012]
Here, the “medical image” refers to a digital image such as an X-ray image or a CT image used for image diagnosis, and can be a digital chest X-ray image or a chest CT image, for example.
[0013]
The “structure normal image” refers to an image having a structure (shape and texture) when the structure of the structure is normal. This means that the “structure normal image” does not include information on other structures other than the predetermined structure. Therefore, a “structure removal image” is a place where a plurality of structures overlap, unlike an image where a structure image has been cut out so that no information remains in the place where the structure image has been removed. In FIG. 5, the image represents a state in which only a predetermined structure is removed.
[0014]
As the “image representing the normal structure of the structure of the same kind as the predetermined structure” and “structure change vector”, for example, a plurality of different images representing the normal structure of the structure of the same kind are sampled (teacher data) The statistical image of the average shape and average texture of the structure is referred to as an “image representing the normal structure of the structure of the same type as the given structure”, and the main component of the shape and texture of the structure A product obtained by multiplying the eigenvector of shape and texture by a coefficient, which is calculated by performing analysis, can be referred to as a “structure change vector”.
[0015]
In the step of detecting an abnormal shadow and the abnormal shadow detecting means, as a method of detecting an abnormal shadow, a method of detecting a tumor shadow using an iris filter or the like proposed in Japanese Patent Laid-Open No. 2002-293354 is used. Can be used.
[0016]
Further, in the first abnormal shadow detection method, the step of generating the structure normal image includes the step of acquiring image data corresponding to an image representing a normal structure of the same kind of subject as the subject, and the same kind of subject. An object representing a normal structure of the object corresponding to the medical image by setting a plurality of structure change vectors for changing the structure of the image and adding the plurality of structure change vectors to the image data A normal image is generated using a step of generating a normal image and a neural network that extracts a normal image of a structure of the same type as the structure in the normal subject image from the normal subject image based on the normal subject image. And a step of generating an image.
[0017]
Similarly, in the first abnormal shadow detection apparatus, the structure normal image generation means includes image data acquisition means for acquiring image data corresponding to an image representing a normal structure of the same type of subject as the subject, and the same type A structure change vector setting means for setting a plurality of structure change vectors for changing the structure of the subject, and by adding the plurality of structure change vectors to the image data, the structure of the subject corresponding to the medical image is added. Based on the subject normal image, a structure change vector adding means for generating a subject normal image representing a normal structure, and a normal image of the same type of structure as the structure in the subject normal image is extracted from the subject normal image. An image generation means for generating the normal structure image can be provided using a neural network.
[0018]
As a method for artificially generating an image as described above, for example, a method based on AAM (Active Appearance Model) can be used. Details of the AAM method will be described later.
[0019]
When an image corresponding to a structure image in a medical image is artificially generated, as described above, the image represents a normal structure of the structure, but this includes an abnormal structure. If the inter-image calculation is performed using images, in addition to the abnormal elements inherent in the medical image, other abnormal elements will enter the resulting image, resulting in an image that is not suitable for image diagnosis. is there.
[0020]
The second abnormal shadow detection method of the present invention is a structure image representing at least one predetermined structure among the plurality of structures in a medical image representing a subject composed of a plurality of structures having anatomical features. Selecting an image that substantially matches the image from a plurality of different images that are artificially generated and representing a normal structure of a structure of the same type as the predetermined structure, and a medical image and the selected image. A step of generating a structure removal image in which the structure image is removed from the medical image by an inter-image calculation, and a step of detecting an abnormal shadow in the structure removal image based on the structure removal image. It is a characteristic method.
[0021]
The second abnormal shadow detection apparatus of the present invention includes an image input means for inputting a medical image representing a subject composed of a plurality of structures having anatomical features, and the plurality of structures artificially generated. Image storage means for storing a plurality of different images representing a normal structure of a structure of the same type as at least one predetermined structure of the objects, and substantially matches the structure image representing the predetermined structure in the medical image A structure-removed image obtained by removing a structure image from a medical image by image selection means for selecting an image from images stored by the image storage means, and an inter-image calculation between the medical image and the selected image. The structure removal image generation means for generating the image and the abnormal shadow detection means for detecting the abnormal shadow in the structure removal image based on the structure removal image are provided.
[0022]
In the first abnormal shadow detection method and apparatus, the structure normal image corresponding to the structure image representing the predetermined structure is artificially generated based on the medical image, whereas the second abnormal shadow detection method is used. The apparatus is different from the apparatus in that an image matching a structure image representing a predetermined structure is selected from a plurality of artificially generated images.
[0023]
In the step of selecting the image and the image selection means, there is a method for searching for a similar image by using a histogram (density signal) of image data as described in JP-A-2002-008350. For example, a method of generating a histogram of density signals for each of the medical images and the stored images, and selecting an image whose histogram is most similar to the medical images can be used. The determination of whether or not they are similar can be made, for example, by calculating a variance value of the difference between the histograms of the medical image and the stored image, and determining that the smaller the variance value, the more similar. Good.
[0024]
The first program of the present invention corresponds to a structure image representing at least one predetermined structure among the plurality of structures in a medical image representing a subject composed of a plurality of structures having anatomical features. The structure image is removed from the medical image by artificially generating a normal structure image representing the normal structure of the predetermined structure and an inter-image calculation between the medical image and the generated normal structure image. Generating the abnormal structure detection image, and detecting the abnormal shadow in the structure removal image based on the structure removal image. It is a program.
[0025]
In addition, the second program of the present invention provides a structure image representing at least one predetermined structure among the plurality of structures in a medical image representing a subject composed of a plurality of structures having anatomical features. Selecting a substantially matching image from a plurality of different images representing a normal structure of a structure of the same type as the predetermined structure, which is artificially generated, and a medical image and the selected image A step of generating a structure removal image in which the structure image is removed from the medical image by an inter-image operation; and a step of detecting an abnormal shadow in the structure removal image based on the structure removal image. Is a program for causing a computer to execute the abnormal shadow detection process.
[0026]
Such first and second programs are recorded on a computer-readable recording medium, and the recording medium is read to be installed in a computer for execution, or on a server connected to a network. It may be stored in a downloadable state and downloaded from the server so as to be installed and executed on a computer.
[0027]
As the “inter-image calculation”, it is particularly preferable to apply a subtraction process in which pixels between image information representing the respective images are associated with each other for the two images. In this case, simple subtraction may be used, or subtraction after weighting may be performed. The difference image obtained by the subtraction process is generally referred to as a subtraction image.
[0028]
In the present invention, the “medical image” may be an image representing the chest. In this case, the “plural structures having anatomical features” may be ribs, clavicles, lungs, bronchi, or the like, or bones representing ribs or clavicles, soft parts representing soft tissues, or the like.
[0029]
Further, if the “predetermined structure” is a bone part, an abnormal shadow can be detected based on a soft part image representing only a soft tissue from which a bone part such as a rib or a clavicle has been removed. If the `` structure '' is an entire structure, an abnormal shadow is detected based on a difference image between the chest image to be diagnosed and an image when the chest is normal, that is, an image in which the abnormal shadow is emphasized. Can do.
[0030]
The above difference image obtained when the “predetermined structure” is the entire structure is an artificially generated image representing the normal structure of the entire subject, and an image between the medical image and the generated image. Although it is similar to the difference image obtained by the inter-operation, even if the size and shape of some structures constituting the subject are the same, the size, shape, and texture of other structures are not necessarily the same. The method of generating an image corresponding to each structure as in the invention is advantageous in that an image with higher accuracy that matches the medical image can be generated.
[0031]
In the present invention, the “image” includes not only an actually displayed image having visibility but also image data for displaying the image.
[0032]
【The invention's effect】
According to the first abnormal shadow detection method and apparatus of the present invention, a structure normal image representing a normal structure of a structure constituting a subject is artificially generated, and an image between a medical image and a structure normal image is generated. An image obtained by removing the structure image from the medical image is obtained by calculation, and the abnormal shadow is detected based on the image. Therefore, it is possible to detect an abnormal shadow that could not be sufficiently removed by a simple filtering process as in the prior art. On the other hand, a structure image as a background feature image having a complicated structure that becomes an obstacle can be extracted and removed according to an image pattern of a medical image, and the detection performance of abnormal shadows can be improved.
[0033]
In the first abnormal shadow detection method, the step of generating a structure normal image is the same as the step of acquiring image data corresponding to an image representing a normal structure of a structure of the same type as the predetermined structure. If a step of setting a plurality of structure change vectors for changing the structure of the structure and a step of adding the plurality of structure change vectors to the image data are set, the structure change vector is set. Thus, an image having a substantially arbitrary shape or texture can be generated, so that a target normal structure image can be generated with high accuracy, and the abnormal shadow detection performance can be further improved.
[0034]
Further, in the first abnormal shadow detection apparatus, the structure normal image generation means includes image data acquisition means for acquiring image data corresponding to an image representing a normal structure of a structure of the same type as the predetermined structure. A structure change vector setting means for setting a plurality of structure change vectors for changing the structure of the same kind of structure, and a structure change vector addition means for adding the plurality of structure change vectors to the image data. If it is provided, an image having a substantially arbitrary shape or texture can be generated by setting the structure change vector, so that the target normal structure image can be generated with high accuracy and abnormal shadows can be generated. The detection performance can be further improved.
[0035]
In the first abnormal shadow detection method, the step of generating a normal structure image includes the step of obtaining image data corresponding to an image representing a normal structure of the same type of subject as the subject, and the structure of the same type of subject. A normal object image representing a normal structure of the subject corresponding to the medical image by adding a plurality of structural change vectors to the image data And using a neural network that extracts a normal image of a structure of the same type as the structure in the normal subject image from the normal subject image based on the normal subject image, If it is made to have a step that has a generation step, the approach by the approach different from the method by the AAM is further increased. Can expect the generation of highly accurate structure normal image, it may be possible to further improve the detection performance of the abnormal shadow.
[0036]
Further, in the first abnormal shadow detection apparatus, the structure normal image generation means includes image data acquisition means for acquiring image data corresponding to an image representing a normal structure of the same kind of subject as the subject, and the same kind of the same kind. The structure change vector setting means for setting a plurality of structure change vectors for changing the structure of the subject, and the normality of the subject corresponding to the medical image by adding the plurality of structure change vectors to the image data A structure change vector addition means for generating a normal subject image representing a normal structure, and a neural network for extracting a normal image of a structure of the same type as the structure in the normal subject image from the normal subject image based on the normal subject image If the image generation means for generating the structure normal image is provided using a net, similarly, the AAM method and By way of different approaches, further we can expect the generation of highly accurate structure normal image, it may be possible to further improve the detection performance of the abnormal shadow.
[0037]
According to the second abnormal shadow detection method and apparatus of the present invention, an image corresponding to an image representing a normal structure of a structure constituting a subject is selected from a plurality of artificially generated images, Since an image obtained by removing the structure image from the medical image is obtained by the inter-image calculation between the medical image and the selected image, and the abnormal shadow is detected based on the image, the first abnormal shadow detection method and apparatus are provided. Similarly, a structure image as a background feature image having a complicated structure that hinders the detection of abnormal shadows, which cannot be sufficiently removed by a simple filtering process as in the past, is used as an image pattern of a medical image. Accordingly, it can be extracted and removed, and the detection performance of abnormal shadows can be improved.
[0038]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described. First, a normal image representing a normal structure of a subject or its structure corresponding to a predetermined image such as the subject or a structure constituting the subject is artificially generated. The method to do is demonstrated.
[0039]
As a technique for artificially generating a normal image, “Active Appearance Model (AAM)” [T. F. Coutes, and C.C. J. et al. Taylor, Active Appearance Models, Proc. European Conference on Computer Vision, Vol. 2, pp. 484-498, Springer, 1998], where a number of sample images representing the normal structure of the subject or its structure, and the anatomical feature shapes (landmarks) extracted from these sample images. Used as teacher data and applied to the AAM.
[0040]
Create a statistical model (shape change model, texture change model, shape-texture correlation model) of the normal structure of the subject or its structure from the teacher data, and use this model to support a given image A normal image having an arbitrary shape and texture can be artificially generated.
[0041]
As the sample image, it is preferable to use an actually captured image. However, if a sufficient number of images cannot be prepared, the sample image is randomly maintained while maintaining the characteristic and normal basic structure of the subject or its structure. The created artificial image may be used.
[0042]
Hereinafter, a procedure for artificially generating a normal image corresponding to a predetermined image such as a subject or a structure thereof will be described taking as an example a case where a normal chest image corresponding to a chest image is generated.
[0043]
FIG. 1 is a flowchart illustrating a normal image generation procedure.
[0044]
(1) Creation of average shape and average texture (step S1)
First, the shape of the anatomical feature of the chest is extracted as a landmark from each sample image representing the chest. As an example, FIG. 2 is a diagram showing a state in which landmarks are attached to the outer edges of the lung field regions for the sample images P1, P2, and P3, and a plurality of black dots represented by symbol M in the figure are landmarks. It is. The points used as landmarks are equivalent anatomical feature positions in each sample image. An average shape of the chest is created from each shape extracted as a landmark in this way. FIG. 3 is a diagram showing an image Pa representing the average shape of the chest. When extracting each shape as a landmark, it is better to normalize the translation component in advance. For example, in the chest image, the upper end of the lung field is aligned with the center in the horizontal direction (left-right direction).
[0045]
Next, in order to warp each sample image to the created average shape, a shift amount of each landmark point in each sample image to the average shape is calculated. FIG. 4 shows how each point of a landmark (represented by symbol M in the figure) attached to the outer edge of the lung field region in the chest image is warped to an average shape (represented by symbol Ka in the figure). It is a figure. Then, the shift amount of each landmark point is approximated using a two-dimensional quintic polynomial. Expressions (1) to (4) show mathematical expressions of the polynomial approximation.
[0046]
[Expression 1]

Here, x and y are the coordinates of each landmark point in each sample image, x ′ and y ′ are the coordinates on the average shape to be warped, Δx and Δy are the shift amounts to the average shape, n is the order, and aij , Bij are coefficients. Note that the coefficient of polynomial approximation is obtained using the method of least squares.
[0047]
Using the above equation, the shift amount to the average shape for each pixel is calculated for each sample image, and each sample image is warped to the average shape for each pixel. At this time, for a pixel whose coordinates after warping move to a position including a decimal point instead of an integer, a pixel value is obtained by first order approximation from four neighborhoods. That is, pixel values are distributed in proportion to the distance from the coordinates after warping to the coordinates of each pixel for the four pixels surrounding the coordinates after warping. Then, an average texture is created from each sample image after warping to the average shape.
[0048]
At this point, an average shape for the chest, each sample image after warping to the average shape, and an average texture are obtained.
[0049]
(2) Creation of shape change model (step S2)
Using the shape of the chest in each sample image and the average shape thereof, a principal component analysis is performed on the shape of the chest, and a unique shape is calculated. Intrinsic components can be found in the literature [Matthew Turk, Alex Pentland, Eigenfaces for Recognition, Journal of Cognitive Neuroscience, vol. 3, num1, 1991]. An arbitrary shape can be approximated by the following equation using a linear sum of eigen shapes.
[0050]
[Expression 2]

Here, x is a shape vector (x1, y1,..., Xi, yi,..., Xn, yn), xave is an average shape vector, Ps is an eigenvector of shape, and bs is a set (combination) of shape coefficients. It is. FIG. 5 shows different unique shapes K1, K2, and K3 as landmarks M as examples of the unique shapes. In the figure, the solid line Ka is an average shape.
[0051]
(3) Creation of texture change model (step S3)
Using the chest texture and its average texture in each sample image, a principal component analysis is performed on the chest texture to calculate a unique texture.
[0052]
[Equation 3]

Here, g is a gray level (texture pixel value) vector (g11, g21,..., Gij,..., Gnn), gave is an average gray level vector, Pg is a gray level eigenvector, and bg is A set of gray level coefficients. FIG. 6 shows different unique textures T1, T2, and T3 as examples of the unique texture.
[0053]
(4) Creation of shape and texture correlation model (step S4)
The shape / texture of the chest in each sample image can be expressed by using coefficients bs and bg composed of values projected on each unique shape / natural texture. As shown below, the principal component analysis is further performed on these coefficients, and appearance parameters for controlling the shape and the gray level are calculated.
[0054]
[Expression 4]

Here, Ws is a diagonal matrix and a weight for each shape factor.
[0055]
[Equation 5]

Q is an eigenvector, and c is an appearance parameter (the average of c is 0) that controls the shape and gray level. By changing the appearance parameter, an almost arbitrary image of the chest can be artificially generated.
[0056]
(5) Generation of normal image (step S5)
A normal chest image corresponding to the chest image P is generated by the following procedure.
1. Appearance parameters calculated from the average shape and images within the average shape are set as initial values (normal image data acquisition means).
2. Create an artificial image by gradually changing the parameters in order from the appearance parameters of the upper eigencomponents, and search for a parameter whose difference between the artificial image and the chest image is less than a predetermined value (structure change vector setting means, structure change) Vector addition means).
3. 2. The artificial image obtained from the value of the appearance parameter obtained in the above is determined as a normal chest image corresponding to the chest image P.
[0057]
According to the above method, if each element such as the average shape / average texture, eigen shape / eigen texture, eigen vector Q of the subject is obtained in advance, a substantially arbitrary image corresponding to the medical image can be obtained only by adjusting the appearance parameter c. Can be generated.
[0058]
In order to generate an image corresponding to a predetermined structure that constitutes a subject, an image that represents only the predetermined structure may be prepared as a sample image. For example, the entire subject actually photographed is represented. An image obtained by extracting only an area representing only the predetermined structure from the image can be generated and used as a sample image. This region extraction can be realized by using an image recognition technique such as a template matching method or a method of detecting an edge by examining a change in shading, or by manually cutting out.
[0059]
(Example 1)
FIG. 7 is a schematic block diagram showing a configuration of a first abnormal shadow detection apparatus which is an embodiment of the first abnormal shadow detection apparatus of the present invention. The first abnormal shadow detection apparatus shown in FIG. 7 has a thoracic radiation image P (hereinafter simply referred to as a chest image) having an anatomical feature and a human chest composed of bones such as ribs and clavicles and other soft parts. And the subject image input means 10 for inputting the image and the bone normal image Sh representing the normal structure of the bone corresponding to the bone image Ph representing the bone in the chest image P, and artificially generated. A structure-removed image generating unit that generates a bone-removed image Rh representing an image in which a bone part has been removed from the chest by an inter-image calculation between the structure normal image-generating unit 20 and the chest image P and the bone normal image Sh. 30, an abnormal shadow detection unit 40 that performs abnormal shadow detection based on the generated bone part removal image Rh, and a display unit 50 that displays the generated bone part removal image Rh on the screen together with the abnormal shadow detection result. ing.
[0060]
Next, the operation of the first abnormal shadow detection apparatus configured as described above will be described. FIG. 8 is a diagram showing a processing flow of the apparatus.
[0061]
The subject image input means 10 inputs a chest image P representing a radiation image of a human chest from a CR device capable of acquiring a radiation image as a digital image, a storage device storing a digital image of the radiation image, or the like (step S11). Here, it is assumed that the chest image P includes an abnormal shadow f representing a cancerous portion behind the bone. FIG. 9 is a diagram showing such a chest image P.
[0062]
The structure normal image generating means 20 obtains the bone normal image Sh by the following processing based on the input chest image P.
[0063]
First, a bone edge, that is, a boundary between a bone part and other soft tissue is detected by detecting a steep change in image density in the chest image P, and the chest image P is represented as a bone part image representing the bone part. Recognition is performed by dividing into Ph and a soft part image Pg representing the soft part (step S12).
[0064]
Next, a normal bone part image Sh representing the normal structure of the bone part corresponding to the recognized bone part image Ph, and a soft part normal image Sg representing the normal structure of the soft part corresponding to the recognized soft part image Pg. Are tentatively generated (step S13), and the bone normal image Sh and the soft normal image Sg are combined to generate an integrated image Shg representing the entire chest (step S14).
[0065]
The normal bone part image Sh and the normal soft part image Sg are obtained by using the above-described AAM technique, using a sample image representing only a normal bone part and a sample image representing only a normal soft part as teacher data. An arithmetic expression for artificially generating an image is obtained, and an appropriate appearance parameter is searched for each arithmetic expression.
[0066]
Then, the generated integrated image Shg and the original chest image P are compared to determine whether or not the image patterns are sufficiently close to each other (step S15). The image Sh is determined as an image corresponding to the bone part image Ph (step S16). Otherwise, the appearance parameter is changed for one or both of the bone normal image Sh and the soft normal image Sg. The bone normal image Sh corresponding to the bone image Ph is obtained by repeating the same process and repeating the same processing. Note that whether or not the image pattern is sufficiently close is determined to be close enough, for example, when the variance of the difference signal value between corresponding pixels between the two images to be compared is within a predetermined value. be able to. FIG. 10 is a diagram showing the bone normal image Sh obtained in this way.
[0067]
The structure removal image generation means 30 performs subtraction processing as an inter-image calculation between the chest image P and the bone normal image Sh obtained by the structure normal image generation means 20, specifically, from the chest image P. By subtracting the bone normal image Sh from corresponding pixels, a bone removed image Rh representing an image in which the bone is removed from the chest is generated (step S17). FIG. 11 is a diagram showing the bone removal image Rh. Since the bone part removal image Rh thus obtained retains the soft part information and the abnormal shadow information originally possessed by the chest image P, it can be used as an image corresponding to the energy subtraction image.
[0068]
The abnormal shadow detection means 40 detects an abnormal shadow based on the generated bone part removal image Rh by using various existing abnormal shadow detection methods (step S18). Examples of the abnormal shadow detection method include a method of detecting a tumor shadow using an iris filter proposed in Japanese Patent Application Laid-Open No. 2002-109510.
[0069]
The display means 50 displays the bone removed image Rh together with the abnormal shadow detection result on a screen such as a CRT or a liquid crystal panel (step S19). The display means 50 may display the normal bone portion image Sh and the original chest image P side by side on the same screen, or may align and display these two images.
[0070]
In the first embodiment, when the bone normal image Sh and the soft part normal image Sg are provisionally generated, only the image having a similar image pattern is selected while being compared with the recognized bone part Ph and the soft part image Pg. The integrated image Shg is generated using them, but the appearance parameter is appropriately set in the calculation formula for generating each image without comparing the recognized bone image Ph and the soft image Sg. The integrated image Shg may be generated each time it is changed. In this case, in the process until the bone normal image Sh corresponding to the bone image Ph is found and fixed, the bone normal image Sh, the soft normal image Sg, and the integrated image Shg are not necessarily generated as image data. Instead, the gray level coefficient set bg for each image may be used as information for identifying these images.
[0071]
Further, in the process of tentatively generating the bone normal image Sh and the soft part normal image Sg and combining them to generate the integrated image Shg, the calculation for generating the bone normal image Sh and the soft part normal The calculation for generating the image Sg may be any process as long as the calculation is substantially different, may be processed separately, or may be processed in an integrated manner. Also good. For example, the gray level u of the soft part and the gray level h of the bone part are respectively expressed by using the above equation (6),
[Formula 6]

The gray level of the integrated image
[Expression 7]

As a matter of course, it may be handled as an integrated process.
[0072]
Further, in the first embodiment, the bone normal image Sh is subtracted from the chest image P to obtain a bone removed image Rh representing an image obtained by removing the bone from the chest. By subtracting the integrated image Shg from, a bone soft part removed image Rhg representing an image obtained by removing the bone part and the soft part from the chest may be obtained. In the bone soft part removal image Rhg obtained in this way, information on the abnormal shadow originally possessed by the chest image P remains, so that it can be used as an image corresponding to the temporal subtraction image.
[0073]
(Example 2)
A second abnormal shadow detection device according to another embodiment of the first abnormal shadow detection device of the present invention will be described. The second abnormal shadow detection device has the same configuration as the first abnormal shadow detection device, and as shown in FIG. 7, the subject image input means 10, the structure normal image generation means 20, and the structure removal image generation means. 30, an abnormal shadow detection unit 40, and a display unit 50, but the normal image generation process in the structure normal image generation unit 20 is different.
[0074]
Next, the operation of the second abnormal shadow detection device will be described. FIG. 12 is a diagram illustrating a processing flow of the apparatus.
[0075]
The subject image input means 10 inputs a chest image P representing a radiation image of a human chest from a CR device, a storage device, or the like (step S21).
[0076]
The structure normal image generating means 20 obtains the bone normal image Sh by the following processing based on the input chest image P.
[0077]
First, the edge of the bone, that is, the boundary between the bone part and the other soft tissue is detected by detecting a steep change in image density in the chest image P, and the soft part image Pg representing the soft part is recognized (step S22). ).
[0078]
Next, a soft part normal image Sg representing the normal structure of the soft part corresponding to the recognized soft part image Pg is generated (step S23), and the soft part normal image Sg is subtracted from the chest image P by corresponding pixels. Thus, a soft part removed image Rg representing an image obtained by removing the soft part from the chest is generated (step S24). At this time, the soft part removal image Rg includes bone part information and abnormal shadow information. Further, a bone edge is detected in the soft part removal image Rg, and a bone part image Ph ′ representing the bone part in the soft part removal image Rg is recognized (step S25). A normal bone part image Sh ′ representing the normal structure of the bone part corresponding to the recognized bone part image Ph ′ is generated (step S26).
[0079]
The normal soft part image Sg and the normal bone part image Sh ′ are obtained by using the above-described AAM technique, using a sample image representing only a normal soft part and a sample image representing only a normal bone part as teacher data. An arithmetic expression for artificially generating a normal image is obtained, and an appropriate appearance parameter is searched for each arithmetic expression.
[0080]
The structure removal image generation means 30 performs subtraction processing as an inter-image operation between the soft part removal image Rg obtained by the structure normal image generation means 20 and the bone normal image Sh ′, specifically, soft part removal. By subtracting the bone normal image Sh ′ from the corresponding pixels from the image Rg, a soft bone removal image Rgh representing an image in which the soft part and the bone part are removed from the chest is generated (step S27). In the soft bone removal image Rgh obtained in this way, information on the abnormal shadow originally possessed by the chest image P remains, so that it can be used as an image corresponding to the temporal subtraction image.
[0081]
The abnormal shadow detection means 40 detects an abnormal shadow by the above-described various abnormal shadow detection methods based on the generated soft bone removal image Rgh (step S28).
[0082]
The display means 50 displays the soft bone removed image Rgh together with the abnormal shadow detection result on a screen such as a CRT or a liquid crystal panel (step S29). The display means 50 displays the normal bone portion image Sh ′ and the original chest image P, the normal soft portion image Sg and the original chest image P, or the soft bone removal image Rgh and the original chest image P on the same screen. May be displayed side by side, or these two images may be aligned and displayed.
[0083]
Example 3
A third abnormal shadow detection device, which is still another embodiment of the first abnormal shadow detection device of the present invention, will be described. The third abnormal shadow detection device has the same configuration as the first abnormal shadow detection device, and as shown in FIG. 7, the subject image input means 10, the structure normal image generation means 20, and the structure removal image generation means. 30, an abnormal shadow detection unit 40, and a display unit 50, but the normal image generation process in the structure normal image generation unit 20 is different.
[0084]
Next, the operation of the third abnormal shadow detection device will be described. FIG. 13 is a diagram illustrating a processing flow of the apparatus.
[0085]
The subject image input means 10 inputs a chest image P representing a radiation image of the human chest from a CR device, a storage device, or the like (step S31).
[0086]
The structure normal image generating means 20 obtains the bone normal image Sh by the following processing based on the input chest image P.
[0087]
First, a normal chest image Sk representing the normal structure of the chest corresponding to the chest image P is generated (step S32). The normal chest image Sk is obtained by using the above-described AAM method to obtain an arithmetic expression for artificially generating a normal chest image using a sample image representing a normal breast as teacher data. Generated by searching for appropriate appearance parameters.
[0088]
Then, using a neural network in which the correspondence relationship between the chest image and the bone image is learned in advance, a bone image having a correspondence relationship with the normal chest image Sk is obtained, and this bone image Ph in the chest image P is obtained. The bone normal image Sh representing the normal structure of the bone corresponding to is set as a normal bone image Sh (step S33). Here, the “correspondence between the chest image and the bone image” means, for example, that the gray level k of the chest and the gray level h of the bone are respectively expressed by using the above equation (6),
[Equation 8]

If it expresses, it can be set as the correspondence of bk and bh. The details of the neural network are described in Japanese Patent Application Laid-Open No. 2000-48187.
[0089]
The structure removal image generation means 30 performs subtraction processing as an inter-image calculation between the chest image P and the bone normal image Sh obtained by the structure normal image generation means 20, specifically, from the chest image P. By subtracting the bone normal image Sh from corresponding pixels, a bone removed image Rh representing an image in which the bone is removed from the chest is generated (step S34). Since the bone part removal image Rh thus obtained retains the soft part information and the abnormal shadow information originally possessed by the chest image P, it can be used as an image corresponding to the energy subtraction image.
[0090]
The abnormal shadow detection means 40 detects the abnormal shadow by the various abnormal shadow detection methods described above based on the generated bone part removal image Rh (step S35).
[0091]
The display means 50 displays the bone removed image Rh together with the abnormal shadow detection result on a screen such as a CRT or a liquid crystal panel (step S36).
[0092]
According to the first to third abnormal shadow detection devices as the first abnormal shadow detection device according to the present invention, the structure normal image representing the normal structure of the structure constituting the subject is artificially generated. The image obtained by removing the structure image from the medical image is obtained by performing the inter-image calculation between the medical image and the normal structure image, and the abnormal shadow is detected based on the image. A structure image as a background feature image having a complicated structure that hinders the detection of abnormal shadows, which could not be sufficiently removed by simple filtering processing, according to the image pattern of the medical image and removed It is possible to improve the detection performance of abnormal shadows.
[0093]
In the first to third embodiments, a method of artificially generating an image is used to obtain a normal image representing a normal structure corresponding to a predetermined image. A method (second abnormal shadow detection method of the present invention) of selecting an image pattern that substantially matches the image pattern from among a large number of images generated at the same time is also conceivable. An embodiment in which this method is used will be described below.
[0094]
(Example 4)
FIG. 14 is a schematic block diagram showing a configuration of a fourth abnormal shadow detection device which is an embodiment of the second abnormal shadow detection device of the present invention. The fourth abnormal shadow detection apparatus shown in FIG. 14 includes a subject image input means 10 for inputting a chest image P, and a plurality of different images that represent artificially generated normal structures of soft and bone portions in the chest. An image storage unit 60 to store, an image selection unit 70 to select an image Ch that substantially matches the bone image Ph in the chest image P from images stored by the image storage unit 60, and a chest image P to be selected Based on the generated bone removal image Rh and the structure removal image generation means 30 for generating the bone removal image Rh representing the image in which the bone has been removed from the chest by the inter-image calculation with the generated image Ch. An abnormal shadow detection unit 40 that performs abnormal shadow detection, and a display unit 50 that displays the generated bone part removal image Rh on the screen together with the abnormal shadow detection result are provided. The fourth abnormal shadow detection device corresponds to the first abnormal shadow detection device in the first embodiment.
[0095]
Next, the operation of the fourth abnormal shadow detection apparatus configured as described above will be described. FIG. 15 is a diagram illustrating a processing flow of the apparatus.
[0096]
The subject image input means 10 inputs a chest image P representing a radiation image of a human chest from a CR device capable of acquiring a radiation image as a digital image, a storage device storing a digital image of the radiation image, or the like (step S41).
[0097]
Based on the input chest image P, the image selection means 70 obtains a normal bone portion image Sh by the following processing.
[0098]
First, a bone edge, that is, a boundary between a bone part and other soft tissue is detected by detecting a steep change in image density in the chest image P, and the chest image P is represented as a bone part image representing the bone part. Recognition is divided into Ph and a soft part image Pg representing a soft part (step S42).
[0099]
Next, a normal bone part image Sh representing the normal structure of the bone part that substantially matches the recognized bone part image Ph, and a normal soft part representing the normal structure of the soft part that substantially matches the recognized soft part image Pg. The image Sg is provisionally selected from a large number of artificial images of the bone and soft parts stored by the image storage means 60 (step S43), and the bone normal image Sh and the soft part normal image Sg are further selected. An integrated image Shg representing the entire chest is generated by combining (step S44).
[0100]
Then, the generated integrated image Shg and the original chest image P are compared to determine whether or not the image patterns are sufficiently close to each other (step S45). The image Sh is determined as an image corresponding to the bone part image Ph (step S46). If not, one or both of the bone normal image Sh and the soft part normal image Sg are selected again (step S43). By repeating the same processing, a normal bone portion image Sh corresponding to the bone portion image Ph is obtained. At this time, the selected artificial image may be used as it is as the normal bone portion image Sh, but an image obtained by warping the selected artificial image into the shape of the recognized bone portion image Sh is used as the normal bone portion image Sh. You may make it use. Note that whether or not the image pattern is sufficiently close is determined to be close enough, for example, when the variance of the difference signal value between corresponding pixels between the two images to be compared is within a predetermined value. be able to. Since the subsequent steps are the same as those in the first embodiment, a description thereof will be omitted.
[0101]
According to the fourth abnormal shadow detection device as an embodiment of the second abnormal shadow detection device of the present invention, an image corresponding to an image representing the normal structure of the structure constituting the subject is artificially generated. An image obtained by removing the structure image from the medical image by performing an inter-image calculation between the medical image and the selected image, and selecting an abnormal shadow based on the image. As in the case of the system to which the first abnormal shadow detection method of the present invention is applied, the detection is complicated, which is an obstacle to the detection of abnormal shadows that could not be sufficiently removed by a simple filtering process as in the prior art. A structure image as a background feature image having a structure can be extracted and removed according to the image pattern of the medical image, and the detection performance of abnormal shadows can be improved.
[0102]
In addition, although the said Example 4 respond | corresponds to the 1st abnormal shadow detection apparatus in Example 1, the thing corresponding to the 2nd, 3rd abnormal shadow detection apparatus in Example 2, 3 is also considered similarly. be able to.
[0103]
In the first to fourth embodiments, the bone part removal image and the soft part bone part removal image are finally generated. However, it is of course possible to generate the soft part removal image in the same manner.
[0104]
In addition, the structure removal image obtained in this way has applications other than abnormal shadow detection.For example, a bone removal image that mainly represents a soft part visualizes the state of aeration, and shows lung function. The soft part removal image that mainly represents the bone part can also be used for bone diagnosis such as quantification of bone density.
[0105]
(Example 5)
The abnormal shadow detection apparatus of the present invention can be used by being connected to a system comprising various apparatuses via a network or the like.
[0106]
FIG. 16 is a configuration diagram showing an outline of an abnormal shadow detection processing system which is an example of such a system. The system shown in FIG. 16 has an input modality 81 such as a CR device, CT device, or MRI device that obtains image data representing a subject image, and a storage device that stores various data including image data obtained by the input modality 81. 82 and the structure removal image data representing the structure removal image obtained by removing the desired structure from the subject image represented by the image data, and the structure removal image based on the generated structure removal image data An abnormal shadow detecting device 83 for detecting abnormal shadows in the image, a display device 84 for outputting various information such as a subject image, a structure-removed image, and an abnormal shadow detection result to a screen, and a film for outputting the information as a hard copy And a printer 85.
[0107]
Next, the operation of the abnormal shadow detection processing system configured as described above will be described.
[0108]
Image data is input from the input modality 81 or the storage device 82 to the abnormal shadow detection processing device 83. Then, the abnormal shadow detection processing device 83 generates structure removal image data representing a structure removal image obtained by removing a desired structure from the subject image based on the input image data, and the generated structure. Based on the object removal image data, an abnormal shadow in the structure removal image is detected. Then, the abnormal shadow detection result is output to the storage device 82, the display device 84, the film printer 85, etc. alone or in combination with the original image data or the structure removal image data. When storing and saving in the storage device 82, the abnormal shadow detection result is stored in association with the original image data and the structure removal image data. At this time, instead of the structure removal image data, information (a parameter at the time of generation) that can reproduce the generation of the structure removal image data may be stored.
[Brief description of the drawings]
FIG. 1 is a flowchart illustrating a normal image generation procedure.
FIG. 2 is a diagram showing a state in which a landmark is attached to the outer edge of a lung field region with respect to a chest sample image.
FIG. 3 is a diagram showing an image Pa representing the average shape of the chest.
FIG. 4 is a diagram showing a state in which each landmark point attached to the outer edge of the lung field in the chest sample image is warped to an average shape;
FIG. 5 is a view showing different unique shapes by landmarks.
FIG. 6 is a diagram showing different unique textures.
FIG. 7 is a block diagram showing the configuration of the first abnormal shadow detection apparatus.
FIG. 8 is a diagram showing a processing flow of the first abnormal shadow detection apparatus;
FIG. 9 is a diagram showing an input chest image P
FIG. 10 is a diagram showing a bone normal image Sh
FIG. 11 is a diagram showing a bone removal image Rh
FIG. 12 is a diagram showing a processing flow of the second abnormal shadow detection apparatus;
FIG. 13 is a diagram showing a processing flow of the third abnormal shadow detection apparatus;
FIG. 14 is a block diagram showing a configuration of a fourth abnormal shadow detection apparatus;
FIG. 15 is a diagram showing a processing flow of the fourth abnormal shadow detection apparatus;
FIG. 16 is a diagram showing the configuration of an abnormal shadow detection processing system
[Explanation of symbols]
10 Subject image input means
20 Structure image generation means
30 Structure removal image generation means
40 Abnormal shadow detection means
50 Display means
60 Image storage means
70 Image selection means
81 Input modality
82 Storage device
83 Abnormal shadow detection device
84 Display device
85 Film printer
Ch Image of the selected bone
f Abnormal shadow
K1 unique shape
K2 unique shape
K3 unique shape
Ka average shape
M landmark
P Input chest image
P1, P2, P3 chest sample images
Pa Average shape image
Ph (Ph ') bone image
Pg soft part image
Sh (Sh ') Normal bone image
Sg Soft part normal image
Shg integrated image
Sk Normal chest image
Rh bone removal image
Rg Soft part removal image
Rgh Soft bone removal image
T1, T2, T3 specific texture

Claims (11)

A normal structure of the predetermined structure corresponding to a structure image representing at least one predetermined structure among the plurality of structures in a medical image representing a subject composed of the plurality of structures having anatomical features. Artificially generating a structure normal image representing the structure;
Generating a structure-removed image in which the structure image is removed from the medical image by performing an inter-image calculation between the medical image and the generated normal structure image;
And detecting an abnormal shadow in the structure removal image based on the structure removal image. Generating the normal structure image comprises:
Obtaining image data corresponding to an image representing a normal structure of a structure of the same kind as the predetermined structure;
Setting a plurality of structure change vectors for changing the structure of the same type of structure;
The abnormal shadow detection method according to claim 1, further comprising: adding the plurality of structure change vectors to the image data. Generating the normal structure image comprises:
Obtaining image data corresponding to an image representing a normal structure of a subject of the same type as the subject;
Setting a plurality of structure change vectors for changing the structure of the same kind of subject;
Generating a subject normal image representing a normal structure of the subject corresponding to the medical image by adding the plurality of structure change vectors to the image data;
Generating the normal structure image based on the normal subject image using a neural network that extracts from the normal subject image a normal image of a structure of the same type as the structure in the normal subject image. The abnormal shadow detection method according to claim 1, wherein: In a medical image representing a subject composed of a plurality of structures having anatomical features, an image that substantially matches a structure image representing at least one predetermined structure among the plurality of structures is artificially generated. Selecting from a plurality of different images representing a normal structure of a structure of the same type as the predetermined structure;
Generating a structure-removed image in which the structure image is removed from the medical image by performing an inter-image calculation between the medical image and the selected image;
And detecting an abnormal shadow in the structure removal image based on the structure removal image. An image input means for inputting a medical image representing a subject composed of a plurality of structures having anatomical features;
A structure for artificially generating a structure normal image representing a normal structure of the predetermined structure corresponding to a structure image representing at least one predetermined structure among the plurality of structures in the medical image Normal object image generation means;
A structure-removed image generating unit that generates a structure-removed image in which the structure image is removed from the medical image by performing an inter-image calculation between the medical image and the generated normal structure image;
An abnormal shadow detection apparatus comprising: an abnormal shadow detection unit that detects an abnormal shadow in the structure removal image based on the structure removal image. The structure normal image generating means,
Image data acquisition means for acquiring image data corresponding to an image representing a normal structure of a structure of the same type as the predetermined structure;
A structure change vector setting means for setting a plurality of structure change vectors for changing the structure of the same type of structure;
6. The abnormal shadow detection apparatus according to claim 5, further comprising: a structure change vector addition unit that adds the plurality of structure change vectors to the image data. The structure normal image generating means,
Image data acquisition means for acquiring image data corresponding to an image representing a normal structure of a subject of the same type as the subject;
Structure change vector setting means for setting a plurality of structure change vectors for changing the structure of the same type of subject;
A structure change vector addition means for generating a subject normal image representing a normal structure of the subject corresponding to the medical image by adding the plurality of structure change vectors to the image data;
Image generating means for generating the normal structure image using a neural network that extracts a normal image of a structure of the same type as the structure in the normal object image from the normal object image based on the normal object image; The abnormal shadow detection apparatus according to claim 5, further comprising: An image input means for inputting a medical image representing a subject composed of a plurality of structures having anatomical features;
Artificially generated image storage means for storing a plurality of different images representing a normal structure of a structure of the same type as at least one predetermined structure among the plurality of structures;
An image selecting means for selecting an image substantially matching the structure image representing the predetermined structure in the medical image from the images stored by the image storing means;
A structure-removed image generating means for generating a structure-removed image in which the structure image is removed from the medical image by performing an inter-image calculation between the medical image and the selected image;
An abnormal shadow detection apparatus comprising: an abnormal shadow detection unit that detects an abnormal shadow in the structure removal image based on the structure removal image. The abnormal shadow detection apparatus according to claim 5, wherein the medical image is an image representing a chest. A normal structure of the predetermined structure corresponding to a structure image representing at least one predetermined structure among the plurality of structures in a medical image representing a subject composed of the plurality of structures having anatomical features. Artificially generating a structure normal image representing the structure;
Generating a structure-removed image in which the structure image is removed from the medical image by performing an inter-image calculation between the medical image and the generated normal structure image;
And a step for detecting an abnormal shadow in the structure removal image based on the structure removal image. In a medical image representing a subject composed of a plurality of structures having anatomical features, an image that substantially matches a structure image representing at least one predetermined structure among the plurality of structures is artificially generated. Selecting from a plurality of different images representing a normal structure of a structure of the same type as the predetermined structure;
Generating a structure-removed image in which the structure image is removed from the medical image by performing an inter-image calculation between the medical image and the selected image;
And a step for detecting an abnormal shadow in the structure removal image based on the structure removal image.

Images