Development of high‐quality artificial intelligence for computer‐aided diagnosis in determining subtypes of colorectal cancer

Abstract

AbstractBackground and AimThere are no previous studies in which computer‐aided diagnosis (CAD) diagnosed colorectal cancer (CRC) subtypes correctly. In this study, we developed an original CAD for the diagnosis of CRC subtypes.MethodsPretraining for the CAD based on ResNet was performed using ImageNet and five open histopathological pretraining image datasets (HiPreD) containing 3 million images. In addition, sparse attention was introduced to improve the CAD compared to other attention networks. One thousand and seventy‐two histopathological images from 29 early CRC cases at Kyoto Prefectural University of Medicine from 2019 to 2022 were collected (857 images for training and validation, 215 images for test). All images were annotated by a qualified histopathologist for segmentation of normal mucosa, adenoma, pure well‐differentiated adenocarcinoma (PWDA), and moderately/poorly differentiated adenocarcinoma (MPDA). Diagnostic ability including dice sufficient coefficient (DSC) and diagnostic accuracy were evaluated.ResultsOur original CAD, named Colon‐seg, with the pretraining of both HiPreD and ImageNET showed a better DSC (88.4%) compared to CAD without both pretraining (76.8%). Regarding the attentional mechanism, Colon‐seg with sparse attention showed a better DSC (88.4%) compared to other attentional mechanisms (dual: 79.7%, ECA: 80.7%, shuffle: 84.7%, SK: 86.9%). In addition, the DSC of Colon‐seg (88.4%) was better than other types of CADs (TransUNet: 84.7%, MultiResUnet: 86.1%, Unet++: 86.7%). The diagnostic accuracy of Colon‐seg for each histopathological type was 94.3% for adenoma, 91.8% for PWDA, and 92.8% for MPDA.ConclusionA deep learning‐based CAD for CRC subtype differentiation was developed with pretraining and fine‐tuning of abundant histopathological images.