A multi-class gastric biopsy artificial intelligence model developed from whole slide histopathological images

Abstract

Pathological diagnosis of biopsies remains the gold standard for gastric diseases such as gastric carcinoma (GC), but it suffers from clinical misdiagnosis and high workload. Accurately distinguishing early-stage GC (EGC) from advanced-stage GC (AGC) and preoperatively predicting lymph node metastasis (LNM) are crucial for precise treatment; however, this remains a challenge in biopsies. To address these challenges, we retrospectively collected 20,711 whole-slide images (WSIs) from 17,086 patients across six centers and prospectively enrolled 3698 WSIs from 2965 patients to develop and validate a gastric biopsy artificial intelligence model (GBAIM), which performed six-class classification on WSIs. GBAIM achieved 96.1% sensitivity and 95.0% specificity in external cohorts, and 93.4% and 99.0% in the prospective cohort, respectively, demonstrating excellent performance and strong generalization ability. GBAIM enhanced the performance of all nine pathologists in auxiliary diagnostic experiments, raising their accuracy by 1.7%–39.0%, while reducing diagnostic time by 29.4%–50.5%. Besides, the fine-tuned GBAIM-T, which aims to differentiate EGC from AGC, achieved an AUC of 0.907 and 0.826 on internal and external testing sets, respectively, and GBAIM-N, which aims to predict LNM, achieved an AUC of 0.814 and 0.706 on internal and external testing sets, respectively. Taken together, GBAIM is a valuable tool in clinical practice.