子宮内膜がんの組織学的タイプを深層学習で予測する病理組織画像ベースのアプローチ

According to the new 2023 International Federation of Gynecology and Obstetrics staging system for endometrial cancer (EC), EC is classified into aggressive and nonaggressive histological types. Accurate diagnosis of the histological type of EC is crucial for optimizing treatment strategies and predicting patient outcomes.

To develop and validate a deep convolutional neural network for predicting nonaggressive versus aggressive histological types from hematoxylin and eosin (H&E)-stained images of EC specimens.

A deep convolutional neural network named EC-AI^HIS was developed to predict the nonaggressive or aggressive histological type from 1187 EC specimens. Its generalizability and clinical utility were assessed across multiple cohorts and benchmarked against pathological diagnoses. Furthermore, correlations between the model's predictions and molecular subtypes of EC were examined.

EC-AI^HIS achieved an AUC of 0.911 (sensitivity 82%, specificity 83%). In fivefold cross-validation, AUCs ranged from 0.865 to 0.909. External validation yielded an AUC of 0.859 (sensitivity 75%, specificity 83%). EC-AI^HIS maintained robustness on images from different scanners and of suboptimal quality. In clinical simulation settings, it showed higher sensitivity than pathologists and improved junior pathologists' diagnostic accuracy. EC-AI^HIS scores were associated with molecular subtypes of EC and showed potential prognostic utility in the p53abn subtype.

EC-AI^HIS is an effective tool that can assist pathologists in classifying EC histological types.