Deep Learning for Classification of Colorectal Polyps on Whole-slide Images.

摘要: Context: Histopathological characterization of colorectal polyps is critical for determining the risk cancer and future rates surveillance patients. However, this a challenging task suffers from significant inter- intra-observer variability. Aims: We built an automatic image analysis method that can accurately classify different types on whole-slide images to help pathologists with diagnosis. Setting Design: Our based deep-learning techniques, which rely numerous levels abstraction data representation have shown state-of-the-art results various tasks. Subjects Methods: covers five common (i.e., hyperplastic, sessile serrated, traditional tubular, tubulovillous/villous) are included in US Multisociety Task Force guidelines assessment surveillance. developed multiple approaches by leveraging dataset 2074 crop images, were annotated domain expert as reference standards. Statistical Analysis: evaluated our independent test set 239 measured standard machine-learning evaluation metrics accuracy, precision, recall, F1 score their 95% confidence intervals. Results: shows residual network architecture achieves best performance classification (overall accuracy: 93.0%, interval: 89.0%–95.9%). Conclusions: reduce cognitive burden improve efficacy histopathological subsequent follow-up recommendations.