Colorectal cancer management increasingly depends on molecular information, but routine access to non-invasive subtype assessment remains limited. CMS4, the mesenchymal consensus molecular subtype, has an aggressive profile linked with stromal infiltration, epithelial-mesenchymal transition, transforming growth factor-β signalling and poorer prognosis. A multicentre retrospective analysis published in Radiology evaluated whether preoperative multiparametric MRI radiomics can identify CMS4 before surgery. The work focused on patients with stage II or III colorectal cancer who underwent abdominal, pelvic or rectal MRI before surgery between January 2015 and April 2020. By combining contrast-enhanced T1-weighted imaging and T2-weighted imaging features, the MRI radiomics CMS4 score captured information related to tumour enhancement and internal heterogeneity within a clinically familiar imaging workflow.
A Non-invasive Route to Subtype Information
CMS4 identification usually depends on molecular testing, immunohistochemistry or pathological image assessment using tumour biopsy or tissue sections. These routes can be difficult to integrate into routine care because of cost, limited biopsy material and loss of tumour integrity after treatment. A preoperative imaging-based method therefore offers a practical route to subtype assessment without requiring additional tissue sampling, while preserving a view of the whole tumour rather than a limited tissue area.
The cohort included 253 patients, most with left-sided disease and stage II or III colorectal cancer. Patients from one centre formed the training and internal test groups, while patients from two other centres formed the external test group. CMS classification came from immunohistochemistry in the primary cohort. Follow-up ran from surgery until recurrence, metastasis or up to 60 months. Key exclusions included unavailable baseline MRI, incomplete follow-up, previous neoadjuvant therapy, other malignancy and poor image quality.
Tumours were segmented across the whole tumour volume on T2-weighted and contrast-enhanced T1-weighted MRI. Image preprocessing helped reduce scanner-related variation before radiomics features were extracted. A machine learning classifier generated sequence-specific CMS4 scores, and these were then combined into a merged MRI radiomics CMS4 score. The merged score provided a continuous probability rather than only a binary result, supporting a more interpretable risk assessment for use alongside clinicopathological information.
Combined MRI Features Improve Prediction
The merged MRI score brought together two complementary imaging signals. Contrast-enhanced T1-weighted imaging contributed information linked with tumour enhancement, while T2-weighted imaging contributed information on internal heterogeneity. This design fits the biological pattern of CMS4, which involves angiogenesis and stromal content, while remaining based on MRI sequences already used in clinical practice. The two sequences therefore offered different but related views of tumour phenotype.
Must Read: Causal AI Improves Colorectal Pathology Robustness
The merged score achieved strong performance in internal and external testing, with AUC values of 0.85 and 0.84. It also performed better than several deep learning comparators. Sequence-specific results showed that contrast-enhanced T1-weighted imaging tended to perform more strongly than T2-weighted imaging, although the combined score gave the most complete performance profile. Decision curve and reclassification analyses also favoured the merged score over sequence-specific scores and a clinical model.
The score also aligned with clinical risk. CMS4 status was associated with nodal involvement, lymphovascular invasion and perineural invasion. In multivariable modelling, the merged MRI score remained independently associated with CMS4 status. It also independently predicted disease-free survival, alongside mucin content, lymphovascular invasion and perineural invasion. Patients classified as CMS4 had a higher risk of recurrence or metastasis, whether subtype was assessed by immunohistochemistry or by the merged imaging score. This risk separation supported the clinical relevance of the imaging-based classification.
Imaging Signal Tracks Stromal Biology
Biological interpretation linked the imaging score with molecular patterns known to characterise CMS4. Bulk RNA sequencing and public single-cell RNA sequencing data connected the merged MRI score with transforming growth factor-β signalling, angiogenesis and epithelial-mesenchymal transition. These pathways support the link between the radiomics signal and the stromal, invasive biology of CMS4 colorectal cancer. The merged score maintained good performance in the biological group, reinforcing the consistency of the imaging signal.
The merged score correlated with cancer-associated fibroblast signatures, including myofibroblastic and extracellular matrix-related signatures. CMS4 tumours also had higher stromal scores and lower tumour purity scores, reinforcing the association between the imaging signal and stromal enrichment. Single-cell data showed that differentially expressed genes were mainly enriched in fibroblasts, with myofibroblastic cancer-associated fibroblasts showing the strongest enrichment among fibroblast subtypes. These patterns linked radiomics features with the tumour microenvironment rather than only with tumour size or location.
Cell population patterns also differed between CMS4 and non-CMS4 tumours. CMS4 tumours had increased myeloid and fibroblast populations, while non-CMS4 tumours had more plasma cells. In an exploratory neoadjuvant subset, patients classified as CMS4 by the merged MRI score had a lower proportion of treatment response than those not classified as CMS4. Misclassification occurred more often in cases with higher pathological nodal stage, intestinal contents, elevated core imaging features and smaller tumour volumes. Probability maps helped visualise image regions contributing to CMS4 classification.
Preoperative multiparametric MRI radiomics offers a non-invasive route for identifying CMS4 colorectal cancer using contrast-enhanced T1-weighted and T2-weighted imaging features. The merged MRI score shows strong discrimination, supports recurrence or metastasis risk stratification and links imaging patterns with stromal biology. Its association with cancer-associated fibroblast activity gives the approach biological plausibility within the CMS4 phenotype. Retrospective design, patient selection and remaining prediction errors limit immediate clinical translation, so prospective validation remains necessary before routine use in colorectal cancer pathways.
Source: Radiology
Image Credit: iStock
References:
Liu Z, Gu W, Liu L et al. (2026) Interpretable MRI-based Multiparametric Radiomics for Preoperative Prediction of CMS4 Colorectal Cancer. Radiology; 319:2.
