Hyun-Jong Jang, Kwangil Yim, Sung Hak Lee
Artificial intelligence (AI) has emerged as a promising tool for inferring molecular biomarkers directly from digitized histopathologic slides. However, the current evidence in gastrointestinal and hepatobiliary cancers remains fragmented across tumor types and biomarker categories. Relevant studies were systematically identified and screened according to the PRISMA 2020 statement. PubMed, Embase, Web of Science Core Collection, Scopus, and IEEE Xplore were searched from 1 January 2015 to the date of search. Eligible studies involved gastrointestinal or hepatobiliary malignancies, used histopathology images, applied AI-based methods, and reported molecular biomarker prediction or inference. A total of 110 studies were included. Most studies focused on colorectal, gastric, liver, and pancreatic cancers, with microsatellite instability, mutation status, molecular subtypes, and tumor mutational burden being the most commonly investigated targets. Model architectures evolved from conventional convolutional neural networks to multiple-instance learning and transformer-based methods. While many studies reported promising predictive performance, direct comparison across studies remained challenging because of substantial heterogeneity in datasets, model architectures, and validation strategies. AI-based molecular biomarker identification from pathologic slides shows substantial promise in gastrointestinal and hepatobiliary cancers, but current evidence is constrained by heterogeneity and limited validation. Standardized, multicenter studies are needed before routine clinical implementation.