Deep learning based on CD3 histological slides for prediction of colon cancer outcome: analysis of three international stage III colon cancer cohorts

Prognostic stratification in stage III colon cancer remains poor, despite treatment advances. Tumor-infiltrating lymphocytes, particularly CD3+ T cells, are potential prognostic markers, but manual assessment is labor-intensive and not robust. This study aimed to develop a deep learning model for automated analysis of CD3-stained histological slides to improve prognostic prediction. A total of 1737 patients from three international cohorts (PETACC08, PRODIGE-13, and HARMONY) were analyzed. The deep learning model (VGG19) identified tumor core (TC) and invasive margin (IM) regions on CD3-stained slides. Features from VGG19 and UNI models were used to cluster patients using hierarchical classification. Prognostic performance was evaluated using disease-free survival (DFS) across training, internal validation, and external validation sets. Deep learning classifiers identified distinct patient clusters with significantly different DFS based on TC and IM. For both IM and TC analysis, patients in the favorable group had a better DFS in all sets (IM: p < 0.001, p = 0.04, p = 0.02; TC: p = 0.002, p = 0.01, p = 0.12, respectively). Combining classifiers enhanced prognostic accuracy in all sets (p < 0.001, p = 0.01, p = 0.06, respectively). The model outperformed traditional clinical variables and CD3 enumeration, which demonstrated variability across cohorts. Automated deep learning analysis of CD3-stained slides enables robust and reproducible prognostic stratification in stage III colon cancer, independently of staining and scanning variations. This approach holds promise for guiding personalized treatment strategies. ClinicalTrials.gov Identifiers: NCT00265811, NCT00995202.