AI-derived tumor microenvironment features and recurrence risk in microsatellite-stable colon cancer after adjuvant chemotherapy.

Abstract

3646 Background: Despite adjuvant chemotherapy, a substantial number of patients with stage II–III microsatellite stable (MSS) colon cancer relapse. While clinicopathologic and circulating tumor DNA (ctDNA) analysis can be used for risk stratification, there is a need for further improvement in risk prediction. Here, we utilized artificial intelligence (AI) applied to hematoxylin and eosin (H&E) slides, which enables efficient and comprehensive assessment of tumor microenvironment (TME). Methods: We retrospectively analyzed TME of high-risk stage II and III colon cancer patients treated with surgery and adjuvant fluoropyrimidine and oxaliplatin chemotherapy. AI-based quantification of tumor area, stromal area, and TME cells (lymphocytes, fibroblasts, macrophages, and endothelial cells) was performed using Lunit SCOPE IO. We developed a generalized linear model with a backward elimination process integrating AI-derived TME features with conventional clinicopathologic variables (T stage, N stage, tumor differentiation, lymphatic invasion, venous invasion, and perineural invasion). The optimal cutoff for distinguishing between high-risk and low-risk patients for relapses was determined using the Youden index derived from the model's receiver operating characteristic curve. We evaluated how the model stratified the risk group in the training cohort and validated it in an independent cohort. Results: In the training cohort (n = 390), high stromal lymphocyte density (HR 0.42, 95% CI 0.26 – 0.68), high tumor-stromal ratio (HR 0.46, 95% CI 0.29 – 0.73), and low stromal area per fibroblast (HR 0.33, 95% CI 0.21 – 0.53) were associated with favorable DFS among the TME features. In the multivariate analysis, high stromal lymphocyte density (adj HR 0.57, 95% CI 0.35 – 0.95) and low stromal area per fibroblast (adj HR 0.58, 95% CI 0.33 – 0.99) showed significant associations with DFS independent of clinicopathological covariates. The final TME-integrated model, incorporating N stage, lymphatic invasion, stromal lymphocyte density, and stromal area per fibroblast, stratified patients into high- and low-risk groups with 3-year DFS rates of 72.3% and 93.8%, respectively (adjusted HR for high risk 3.29, 95% CI 1.98 – 5.46). The model identified high-risk patients irrespective of clinicopathological risk features. For example, high-risk status was associated with worse outcomes among patients with N2 disease (adjusted HR 6.57, 95% CI 1.58 – 27.34), and also among patients with N0 or N1 disease (adjusted HR 2.49, 95% CI 1.27 – 4.87). The model was validated in an external independent cohort (n = 260; adjusted HR for high-risk 3.68, 95% CI 1.60 – 8.48). Conclusions: TME-integrated model combining AI-derived TME features with clinicopathologic factors significantly improved recurrence risk stratification in MSS colon cancer receiving adjuvant chemotherapy.