Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Automated acquisition of explainable knowledge from unannotated histopathology images

harunashi
July 02, 2021

Automated acquisition of explainable knowledge from unannotated histopathology images

harunashi

July 02, 2021
Tweet

More Decks by harunashi

Other Decks in Science

Transcript

  1. お品書き • Abstract • Introduction • Method • Result •

    おまけ:オートエンコーダーで教師なし分類実装
  2. Key feature generation method • でっかいサイズの病理画像を100個の特徴量まで圧縮したい • STEP1 弱拡⼤の画像に対して •

    STEP2 強拡⼤の画像に対して • STEP3 STEP1で得た結果をSTEP2の結果と⽐較して補完 →100個の特徴量を得る
  3. Key feature generation method score 100 clusters using the ratio

    of the number of positive/negative images based on the similarity to each cluster. k番⽬のクラスタが再発に働くのか⾮再発に働くのか の程度のスコアを付ける。
  4. Key feature generation method ←The 1568 intermediate-layer features were given

    scores uʼi,j,jʼ,kʼ based on the intensity values vʼi,j,jʼ of each node. Again, we used the following simple scoring method:
  5. Key feature generation method • STEP3 • 各⼩画像は STEP1の特徴kについてのIi,j と

    STEP2の特徴kʼについての Iʼi,j を持つ。 • Ii,j , Iʼi,j を0.5を閾値としてposi,negaとし、⼀致しない場合は 解析に使わない • STEP1で得た特徴量の数を合計して予測に⽤いた。
  6. Explainable features from histopathology images • k-meansで得られる100個のセントロイド →100個の特徴、 セントロイドに最も近い画像が特徴を代表する画像 •

    各特徴(セントロイド)が悪性なのか、良性なのかは再発データ に基づいてスコアがつけられている • 代表する画像を病理医が⾒て意味を解釈する
  7. Explainable features from histopathology images • a-j 異常な構造 • c

    癌細胞を含まない間質成分の密集 • g 出⾎ 病理医のコメント Cancers show Gleason patterns 4 or 5 indicating aggressive clinical behavior. Stromal component without cancer cells tends to show dense cellularity compared to those of normal structure.
  8. Explainable features from histopathology images • p グリソンスコア3に相当 • k-o,

    q-s 癌細胞を含まない緩い間質成分 • t 癌細胞を含まない外科的マージン 病理医のコメント Cancers show Gleason pattern 3 indicating indolent clinical behavior. Stromal component without cancer cells tends to show relatively loose cellularity suggesting normal peripheral zone structure. Cauterized extraprostatic connective tissue without cancer cells, which indicate that the surgical margin is free from cancer.
  9. Discussion • The Gleason score is a unique pathological grading

    system, purely based on architectural disorders, without considering cytological atypia. In this study, none of the cancer cells in the images identified by the deep neural networks as representative of high-grade cancer showed severe nuclear atypia or prominent nucleoli. Our results indicate that the central ideas behind Gleasonʼs grading system are sound.
  10. Discussion • Interestingly, representative images of the features nominated by

    the deep neural networks comprised of not only human-established findings but also previously unspotlighted or neglected features of stroma at the noncancerous area.