Interpretable MRI-based Multiparametric Radiomics for Preoperative Prediction of CMS4 Colorectal Cancer

Background Consensus molecular subtypes (CMSs) are associated with prognostic and clinical outcomes in colorectal cancer (CRC); however, effective noninvasive methods for CMS4 identification are limited. Purpose To evaluate whether a radiomics-based machine learning approach can predict CMS4 status in CRC and to explore its biologic relevance and interpretability of radiomics features. Materials and Methods This multicenter, retrospective study included patients with CRC who underwent abdominal, pelvic, or rectal MRI (January 2015 to April 2020). Patients were followed until recurrence or metastasis or up to 60 months, whichever occurred first. A subgroup of patients was randomly divided into a training and an internal test set, whereas another subgroup constituted the external test set. Pathologic tissue and MRI data were collected, including T2-weighted imaging (T2WI) and contrast-enhanced (CE) T1-weighted imaging (T1WI). CMS classification was determined using immunohistochemistry. A machine learning model was developed to generate an MRI radiomics CMS4 score (MRC4s) for predicting CMS4 status. Deep learning models (ResNet50, VGG16, and DenseNet201) were also implemented as comparators. Performance was evaluated using receiver operating characteristic curve analysis. Bulk and single-cell RNA sequencing data from patients with CRC was used to investigate the association between MRC4s and biologic pathways. Results This study included 253 patients (median age, 63 years; IQR, 55-69 years; 163 men). The merged MRC4s, combining CE T1WI and T2WI features, achieved areas under the receiver operating characteristic curve (AUCs) of 0.85 (95% CI: 0.63, 1.00) in the internal and 0.84 (95% CI: 0.73, 0.95) in the external test sets, outperforming state-of-the-art deep learning models (AUC range, 0.70-0.75; all P < .01). Merged MRC4s stratified the risk of recurrent metastasis (hazard ratio, 5.96; P < .001). Transcriptomic analyses revealed the merged MRC4s were associated with transforming growth factor-β and epithelial-mesenchymal transition pathways. Conclusion A machine learning radiomics model based on preoperative multiparametric MRI predicted CMS4 of CRC from other subtypes with strong performance and biologic interpretability supported by transcriptomic analyses. © The Author(s) 2026. Published by the Radiological Society of North America under a CC BY 4.0 license. Supplemental material is available for this article. See also the editorial by Arita and Kim in this issue.