Establishing a method for Automatic identification of Cervical Cancer Cells by a Deep Neural Network

2022 Fiscal Year Final Research Report

• Project/Area Number: 21K21320
• Research Category: Grant-in-Aid for Research Activity Start-up
• Allocation Type: Multi-year Fund
• Review Section: 1002:Human informatics, applied informatics and related fields
• Research Institution: Kyorin University
• Principal Investigator: Tsuboshita Yukihiro 杏林大学, その他部局等, 教授 (50910634)
• Project Period (FY): 2021-08-30 – 2023-03-31
• Keywords: 深層学習 / 機械学習 / 自動診断 / 子宮頸がん検診 / 細胞診

Outline of Final Research Achievements

In the present study, we aimed to realize automatic cancer cell identification that surpasses the identification performance of specialist technicians, and therefore we performed experiments by applying a deep learning to cytological specimens of the uterine cervix. First, cytological specimens were photographed with a digital camera, and a large-scale data set with the presence and degree of atypical cells was created. The number of images in the database was about 2,619, and the number of patients was 550. Next, using this dataset, we conducted a prediction experiment using deep learning. ResNet, MobileNet, and DenseNet were used for the experiment. The best-performing model was DenseNet with the accuracy of 0.90, the sensitivity of 0.87, and the specificity of 0.94, surpassing the performance of morphological examinations performed by specialist technicians (sensitivity of 0.7, specificity of 0.9).

Free Research Field

機械学習

Academic Significance and Societal Importance of the Research Achievements

細胞診画像のデータベースの作成を行い、最終的に患者数550人、画像数2,619枚の医療分野としては大規模なデータベースを作成した。これらの画像には、細胞診の診断結果のみではなく、HPV検査結果、組織診の検査結果も付与されている点で、これまでの細胞診とは一線を画している。さらに、深層学習を用いた分類実験を行い、一番性能が良かったDenseNet121を用いた実験では、正確度0.895、感度0.870、特異度0.943という結果を得た。これは、専門技師の認識性能を上回る結果となり、今後の実用化に向けて強固な橋頭保を得ることができた。