GastritisMIL: An interpretable deep learning model for the comprehensive histological assessment of chronic gastritis.

Abstract

The comprehensive histological assessment of chronic gastritis is imperative for guiding endoscopic follow-up strategies and surveillance of early-stage gastric cancer, yet rapid and objective assessment remains challenging in clinical workflows. We propose a powerful deep learning model (GastritisMIL) to effectively identify pathological alterations on H&E-stained biopsy slides, thereby expediting pathologists' evaluation and improving decision-making regarding follow-up intervals. We have trained and tested GastritisMIL by using retrospective data from 2,744 patients and evaluated discriminative performance across three medical centers (467 patients). GastritisMIL attained areas under the receiver operating curve greater than 0.971 in four tasks (inflammation, activity, atrophy, and intestinal metaplasia) and superior performance comparable to that of two senior pathologists. Specifically, interpretable attention heatmaps generated by GastritisMIL effectively assist junior pathologists in locating suspicious lesion regions across the entire field and minimizing missed diagnosis risk. Moreover, the high generalizability of this developed model across multiple external cohorts demonstrates its potential translational value.