Automating tumor-stroma ratio quantification in colon cancer patients from the UNITED study.

[BACKGROUND] Colon cancer is a leading cause of cancer-related death worldwide, making accurate prognostic tools crucial. The tumor-stroma ratio (TSR) quantifies the proportion of stromal to tumor epithelial tissue and is a prognosticator in colon cancer. Stroma-high tumors are associated with worse outcomes and less benefit from adjuvant chemotherapy (ACT). Transitioning from visual to automated TSR scoring coincides with digital pathology advancements.

[MATERIALS AND METHODS] In this study, we validated a fully automated artificial intelligence-based TSR quantification algorithm in hematoxylin-eosin-stained slides from the UNITED cohort, including 853 stage II and III colon cancer patients who had not received neoadjuvant therapy. The algorithm segmented 11 tissue classes on whole-slide images and calculated TSR as stroma/(stroma + epithelial tumor). A 1-mm region of interest was implemented, and an optimal automated cut-off of 77% was derived using receiver operating characteristic analyses for disease-free survival (DFS) and overall survival (OS).

[RESULTS] The median patient age was 67 years. Stroma-high patients had significantly worse DFS than stroma-low patients (3-year DFS 71% versus 82%, P < 0.001). This prognostic effect remained significant in the multivariate analysis (P = 0.016) and was consistent across stage II and III subgroups. Importantly, worse DFS was observed in both stroma-high stage II and III patients despite ACT (3-year DFS 74% versus 89%, P = 0.015; and 65% versus 79%, P = 0.002, respectively), suggesting potential predictive value. OS was worse in stroma-high patients as well (5-year OS 69% versus 85%, P = 0.002).

[CONCLUSION] We validated a clinically applicable tool for fully automated TSR quantification in colon cancer, including an optimal region of interest and cut-off value for automated scoring. We confirm the independent prognostic value of automated TSR for DFS and OS and highlight its influence on potential ACT benefit, underlining its clinical relevance. Our findings support its integration into digital pathology workflows.