Evaluation of a Deep Neural Network for Automated Classification of Colorectal Polyps on Histopathologic Slides.

摘要: Importance Histologic classification of colorectal polyps plays a critical role in screening for cancer and care affected patients. An accurate automated algorithm the on digitized histopathologic slides could benefit practitioners Objective To evaluate performance generalizability deep neural network polyp slide images using multi-institutional data set. Design, Setting, Participants This prognostic study used collected from January 1, 2016, to June 31, Dartmouth-Hitchcock Medical Center, Lebanon, New Hampshire, with 326 training, 157 an internal set, 25 validation For external 238 179 distinct patients were obtained 24 institutions across 13 US states. Data analysis was performed April 9 November 23, 2019. Main Outcomes Measures Accuracy, sensitivity, specificity model classify 4 major types: tubular adenoma, tubulovillous or villous hyperplastic polyp, sessile serrated adenoma. Performance compared that local pathologists’ at point identified corresponding pathology laboratories. Results evaluation ground truth labels 5 pathologists, had mean accuracy 93.5% (95% CI, 89.6%-97.4%) 91.4% 87.0%-95.8%). On test set achieved 87.0% 82.7%-91.3%), which comparable 86.6% 82.3%-90.9%). Conclusions Relevance The findings suggest this may assist pathologists by improving diagnostic efficiency, reproducibility, screenings.