AI-Assisted Risk Stratification in Stage II Colorectal Cancer: Multi-Institutional Validation of Semantically-Enhanced Deep Learning
- Author(s)
- Magisson, F; He, Z; Millward, J; Harris, A; Chen, Z; Mielke, LA; Tran, K; Ward, RL; Hawkins, NJ; Sieber, OM; Wong, R; Shapiro, J; Harris, S; Khattak, A; Burge, M; Gibbs, P; Tie, J; Williams, DS;
- Journal Title
- Gastroenterology
- Publication Type
- Jul 28
- Abstract
- BACKGROUND AND AIMS: Accurate risk stratification in stage II colorectal cancer is essential for treatment decision-making, as current guidelines recommend adjuvant chemotherapy only for patients with a high-risk of relapse. We aimed to develop and validate an AI-based approach for automated invasive front assessment to improve prognostic stratification in this population. METHODS: We developed SÉMIL (Semantically-Enhanced Multiple Instance Learning), integrating vision-language foundation models with attention-based multiple instance learning for automated invasiveness assessment from H&E-stained whole slide images. We trained and validated SÉMIL on 1,608 H&E-stained WSIs from three cohorts (Austin n=697, MCO n=478, DYNAMIC n=433). We compared SÉMIL performance against manual pathologist assessment and non-semantic MIL approaches. RESULTS: For binary classification, SÉMIL outperformed non-semantic MIL across all cohorts (external validation: AUC 0.713-0.821 vs 0.686-0.803). For survival prediction, SÉMIL demonstrated validated prognostic stratification in both the internal (Austin: HR=4.73, p=0.0012) and the two external (MCO: HR=2.84, p=0.0032; DYNAMIC: HR=2.10, p=0.0396) stage II validation cohorts. Critically, among National Comprehensive Cancer Network (NCCN) guideline-defined high-risk stage II patients, SÉMIL successfully stratified outcomes across all three cohorts (HRs 2.96-3.50, all p<0.05), demonstrating consistent reproducible performance. In multivariate analysis of the combined stage II cohort (n=1,220), SÉMIL retained independent prognostic significance (HR=1.98, p=0.005) after adjusting for conventional clinicopathological features including T stage, MMR status, and lymph node examination adequacy. Concordance analysis between SÉMIL and manual assessment showed concordant infiltrative classification identified the highest-risk group (HR=3.96, p<0.0001), with discordant cases showing intermediate risk. CONCLUSIONS: SÉMIL demonstrates validated prognostic stratification in stage II colorectal cancer, with potential utility for refining risk assessment within NCCN guideline-defined high-risk categories where treatment decisions are most challenging.
- Publisher
- Elsevier
- Keywords
- artificial intelligence; colorectal cancer; multiple instance learning; risk stratification; stage II
- Research Division(s)
- Personalised Oncology
- PubMed ID
- 42521098
- Publisher's Version
- https://doi.org/10.1053/j.gastro.2026.07.009
- Open Access at Publisher's Site
https://doi.org/10.1053/j.gastro.2026.07.009- Terms of Use/Rights Notice
- Refer to copyright notice on published article.
Creation Date: 2026-07-30 09:25:12
Last Modified: 2026-07-30 09:25:28