Construction and exploratory validation of a digital pathology prognostic model for stage II–III colorectal cancer based on contrastive learning and attention mechanisms

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Introduction Although TNM staging is the standard framework for prognostic assessment in colorectal cancer (CRC), patients within the same stage can have heterogeneous outcomes. This study aimed to extract tumor-tissue features from whole-slide images (WSIs), develop an image-derived prognostic model for stage II–III CRC, and conduct an exploratory evaluation of its performance. Methods A transfer-learning tissue classifier was used to identify tumor regions in CRC WSIs. Self-supervised contrastive learning was used to extract tumor features, which were summarized by patient-level K-Means clustering. An attention-based survival model generated a locked image-derived Risk Score. The actual analyzed population comprised 335 patients with locally advanced CRC: 318 patients recorded as stage III and 17 patients with T4N0 disease (high-risk stage II by AJCC eighth-edition anatomic staging). The 265-patient TCGA cohort was randomly divided 8:2 into a training set and an internal holdout set; 70 SAHSYU patients formed a single-center exploratory cohort. Discrimination was quantified with patient-level bootstrap confidence intervals. Cox regression, Kaplan–Meier analysis, one-year calibration, and exploratory decision-curve analysis were performed. A sensitivity analysis excluded all N0 patients without retraining or changing the locked Risk Score. Results The C-index was 0.704 (95% bootstrap CI 0.637–0.768) in training, 0.663 (0.544–0.776) in the internal holdout set, and 0.706 (0.533–0.859) in SAHSYU. Per one training-set standard deviation, the Risk Score was associated with PFS internally (HR 1.688, 95% CI 1.080–2.640; Wald p=0.022) and exploratorily in SAHSYU (HR 6.010, 95% CI 1.170–30.877; p=0.032). In the sensitivity analysis excluding all N0 records without retraining, the C-index remained 0.647 internally and 0.705 in SAHSYU; fixed-threshold log-rank p values were 0.014 and 0.027, respectively. Conclusion The image-derived Risk Score showed preliminary PFS discrimination and risk-stratification potential in stage II–III CRC. These findings require validation in larger multicenter cohorts with more events, longer follow-up, formal calibration, and clinical-utility assessment before treatment decisions can be informed.