Tomita Naofumi ; Abdollahi Behnaz ; Wei Jason ; Ren Bing ; Suriawinata Arief ; Hassanpour Saeed

Abstract

Deep learning-based methods, such as the sliding window approach forcropped-image classification and heuristic aggregation for whole-slideinference, for analyzing histological patterns in high-resolution microscopyimages have shown promising results. These approaches, however, require alaborious annotation process and are fragmented. This diagnostic studycollected deidentified high-resolution histological images (N = 379) fortraining a new model composed of a convolutional neural network and agrid-based attention network, trainable without region-of-interest annotations.Histological images of patients who underwent endoscopic esophagus andgastroesophageal junction mucosal biopsy between January 1, 2016, and December31, 2018, at Dartmouth-Hitchcock Medical Center (Lebanon, New Hampshire) werecollected. The method achieved a mean accuracy of 0.83 in classifying 123 testimages. These results were comparable with or better than the performance fromthe current state-of-the-art sliding window approach, which was trained withregions of interest. Results of this study suggest that the proposedattention-based deep neural network framework for Barrett esophagus andesophageal adenocarcinoma detection is important because it is based solely ontissue-level annotations, unlike existing methods that are based on regions ofinterest. This new model is expected to open avenues for applying deep learningto digital pathology.