Document resource
Background Consensus molecular subtyping (CMS) defines the transcriptomic taxonomy of colorectal cancer (CRC) and guides precision therapy. However, current methods for CMS determination rely primarily on molecular or histopathological analyses of surgical specimens, limiting their preoperative applicability. This study aims to leverage artificial intelligence (AI) to infer CMS directly from routine preoperative computed tomography (CT) images, providing a noninvasive and cost-effective approach for precision oncology in CRC.Methods 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 (SYSU-CRC, n = 416), an internal validation cohort (SYSU-SAH, n = 1,671), and an external validation cohort (Liaoning, n = 357). A deep learning framework, CT4CMS, was developed to predict CMS from preoperative CT images. To enable 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 using five-fold cross-validation in the discovery cohort and verified on the two validation cohorts.Results CT4CMS demonstrated strong classification performance, achieving a cross-validation AUC of 0.867 in the discovery cohort. In both validation cohorts, patients predicted as CMS4 exhibited significantly poorer disease-free survival (DFS) (SYSU-SAH, P = 3.69 × 10-6; Liaoning, P = 0.043), consistent with transcriptome-defined subtyping trends observed in the discovery cohort (SYSU-CRC, P = 0.025). Importantly, CMS4 patients, particularly those with stage II or III disease, derived substantial benefit from adjuvant chemotherapy compared with surgery alone (SYSU-CRC, P = 3.68 × 10-5; SYSU-SAH, P = 1.36 × 10-4; Liaoning, P = 0.016). Interpretability analysis further revealed subtype-specific radiomic features (all P < 0.05), suggesting biological associations between CT-derived imaging and molecular characteristics.Conclusions This study establishes an interpretable CT-based deep learning framework, CT4CMS, for noninvasive and preoperative CMS prediction in CRC, paving the way for imaging-based molecular stratification and personalized therapeutic decision-making.