Abstract Introduction: The tumor-stroma ratio (TSR), defined as the proportion of stromal content within the primary tumor, has been proposed as a prognostic histopathologic marker for colorectal cancer (CRC). Higher stromal content is often linked to poorer prognosis in stage II-III disease. However, TSR assessment remains manual and is prone to interobserver variability, while existing AI algorithms - often trained and tested on single-institution datasets - lack independent validation for generalizability. In this study, we developed an AI-driven TSR quantification model trained on the public Cancer Genome Atlas (TCGA) pathologist-annotated regions and validated it using predictions on (i) TCGA whole-slide images (WSI) and (ii) two large independent CRC cohorts (NHS/HPFS tissue microarray (TMA)). Methods: We trained a SegFormer semantic segmentation model on 99,871 image tiles derived from pathologist-annotated regions (J.V.) from 469 TCGA H and range from 11.72% to 99.87% with a median of 70.30% for the TCGA cohort. In the NHS/HPFS cohort, AI-derived TSR-high group showed significantly worse outcomes for overall survival (HR = 1.38; 95% CI = 1.06-1.81; P = 0.017) and CRC-specific survival (HR = 1.49; 95% CI = 1.03-2.15; P = 0.035). However, in the TCGA cohort, TSR stratification showed no significant prognostic value for either pathologist-annotated or WSI-AI-derived TSR for overall survival (HR = 1.12, 95% CI = 0.69-1.82; P = 0.647, and HR = 0.97; 95% CI = 0.59-1.60; P = 0.909, respectively) and CRC-specific survival (HR = 0.79, 95% CI = 0.40-1.58; P = 0.513, and HR = 0.88; 95% CI = 0.43-1.77; P = 0.723, respectively). Discussion: While AI-derived TSR showed prognostic significance in the NHS/HPFS cohort, the lack of significance in TCGA suggests limited generalizability across cohorts. Notably, this is, to our knowledge, the first study to evaluate TSR using the well-established TCGA CRC dataset. These findings highlight the need to further dissect the stromal composition - through molecular staining or AI-driven profiling of immune populations, stromal subtypes, and tumor-immune spatial interactions - to capture the biological mechanisms underlying TSR and improve its robustness as a prognostic biomarker. Citation Format: Wei Kit Tan, Marcia Zhang, Juha P. Väyrynen, Shuji Ogino, Mai Chan Lau. Scalable AI-driven tumor-stroma ratio quantification for prognostic stratification in stage II-III colorectal cancer abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 3938.
Tan et al. (Fri,) studied this question.