CT4CMS: Preoperative Computed Tomography-Based Consensus Molecular Subtyping Prediction in Colorectal Cancer Using Interpretable Deep Learning
Authors
Affiliations (1)
Affiliations (1)
- Department of Surgery, The Chinese University of Hong Kong
Abstract
Consensus molecular subtyping (CMS) defines the transcriptomic taxonomy of colorectal cancer (CRC) and guides precision therapy. Although current approaches can predict CMS from histopathology, they rely on surgical specimens, limiting their preoperative applicability. In this study, we developed a deep learning model to infer CMS directly from preoperative computed tomography (CT) scans, enabling noninvasive molecular stratification of CRC. A multi-institutional cohort of 2,444 CRC patients was collected from the Sixth Affiliated Hospital of Sun Yat-sen University and Liaoning Cancer Hospital, comprising a discovery cohort (n = 416), an internal validation cohort (n = 1,671), and an external validation cohort (n = 357). To achieve robust feature extraction, a self-supervised 3D representation learning network was first pretrained on large-scale public CT datasets to capture generalizable imaging features. These representations were subsequently integrated into a multi-instance learning (MIL) classifier for CMS prediction, with attention mechanisms to enhance interpretability. Model performance was evaluated by cross-validation on the discovery cohort and verified on the two validation cohorts. CT4CMS demonstrated strong performance in predicting CMS subtypes directly from CT scans, achieving a cross-validation AUC of 0.867. In both validation cohorts, patients predicted as CMS4 exhibited significantly poorer disease-free survival yet derived substantial benefit from adjuvant chemotherapy, consistent with transcriptome-defined subtyping trends observed in the discovery cohort. Interpretability analysis revealed distinct subtype-specific radiomic features, suggesting that CT-derived imaging features capture underlying molecular characteristics and enable CMS classification. Overall, this study establishes a noninvasive and interpretable deep learning framework for CMS prediction in CRC, paving the way for imaging-based molecular stratification and personalized therapeutic decision-making.