Development and validation of a deep learning–based pathomics signature for prognosis and chemotherapy benefits in colorectal cancer: A retrospective multicenter cohort study.

• DOI: 10.1200/jco.2025.43.4_suppl.298
• Publication Date: 2025-01-27T14:33:47Z
• Authors: Shenghan Lou; Fenqi Du; Huiying Li; Yanming Huang; Laishou Yang; Gen Li; Wenjie Song; Jingmin Xue; Hao Li; Genshen Mo; Hang Wang; Pai Wang; Zhuozhong Wang; Peng Han

298 Background: The current TNM staging system fails to provide adequate information for prognosis and adjuvant chemotherapy benefits in colorectal cancer (CRC). Pathomics, an emerging field, shows promise improving estimation decision-making. In this study, we developed validated a pathomics signature (PS CRC ) that directly analyzes hematoxylin eosin–stained slides using deep learning predict outcomes. Methods: A total of 883 whole slide images from two cohorts, Harbin Medical University Cancer Hospital Genome Atlas (TCGA), were retrospectively analyzed. An interpretable, multi-instance model was proposed establish the PS . Shapley additive explanations employed interpret model's decisions, gradient-weighted class activation mapping applied visualise pathological phenotypes transcriptomics data TCGA cohort used explore potential pathogenesis underlying Results: identified as independent prognostic factor associated with both overall survival disease-free survival. Incorporating into stage resulted significant improvement estimation, evidenced by notable increase net reclassification integrated discrimination improvement. Moreover, among II III patients low levels , satisfactory observed. Notably, main features include tumor cell infiltration, adipocyte accumulation, fibrous tissue deposition, stromal infiltration. Transcriptome analysis further support relevance progression immune suppression. Conclusions: Our finds highlight histopathology images-based predicting assess therapeutic response CRC. could serve effective tool clinical decision management, providing insights pathogenic mechanisms. However, prospective studies are still necessary validation.