Identifying Poor-Prognosis Colorectal Cancer Using MRI Alone — Development of a Preoperative Diagnostic AI Reducing Burden on Patients and Healthcare Costs

A research team led by Dr. Wenchao Gu, specially appointed associate professor at Chiba University Graduate School of Medicine, and Professor Tong Tong of Fudan University Cancer Hospital (Fudan University Cancer Center), has developed an AI-based method to identify particularly aggressive and recurrent-prone 'poor-prognosis colorectal cancer' from MRI images before surgery. Testing on data from 253 colorectal cancer patients across three hospitals confirmed high accuracy in distinguishing these cases, with patients labeled 'high-risk' by the AI showing significantly higher actual recurrence rates. This technology is expected to enable personalized treatment planning prior to surgery.

This research was published in the academic journal Radiology on May 19, 2026.

(Paper link: 10.1148/radiol.251719)

Figure: Overview of the study

Background

Colorectal cancer varies significantly among patients in tumor characteristics and treatment response. The Consensus Molecular Subtypes (CMS) classification divides colorectal cancer into four subtypes, with CMS4 (Note 1) being the most aggressive due to activation of EMT (Note 2) and the TGF-β signaling pathway (Note 3), leading to resistance to chemotherapy and immunotherapy and the poorest prognosis. However, identifying CMS4 requires costly immunohistochemical staining or gene expression analysis, which depend on limited tissue samples obtained post-surgery, making preoperative prediction difficult. To address this, the research team developed a method to predict CMS4 using widely available preoperative MRI images, reducing patient burden and costs.

Key Research Findings

1. Development and validation across multiple institutions: A multiparametric MRI radiomics model named 'MRC4s' was developed using T2-weighted imaging (T2WI) and contrast-enhanced T1-weighted imaging (CE-T1WI) (Note 4).

2. High accuracy surpassing existing deep learning models: The integrated final model (Merge-MRC4s), combining T2WI and CE-T1WI, achieved AUCs of 0.85 in internal validation and 0.84 in external validation, significantly outperforming established deep learning models such as ResNet50, VGG16, and DenseNet201 (AUC 0.70–0.75).

3. Stratification of recurrence risk: Kaplan-Meier analysis showed that patients with high MRC4s scores had approximately six times higher risk of recurrence and metastasis compared to those with low scores, confirming the model’s clinical utility as a prognostic tool.

4. Biological interpretability: The CMS4 group predicted by the model showed significant activation of the TGF-β signaling and EMT pathways. Furthermore, SHAP (Note 5) analysis visualized the contribution of each radiomic feature to predictions, achieving an interpretable AI system.

Dr. Wenchao Gu, Specially Appointed Associate Professor, Chiba University Graduate School of Medicine

Future Outlook (Researcher Comments)

This breakthrough opens the possibility of identifying treatment-resistant CMS4 colorectal cancer using only preoperative MRI. The team plans to expand the model’s application to predict response to neoadjuvant chemoradiotherapy (nCRT) (Note 6) and to select patients for novel therapies targeting CMS4, while advancing validation through prospective multicenter collaborative studies toward clinical implementation.

Glossary

Note 1) CMS4 (Consensus Molecular Subtype 4): One of four molecular subtypes of colorectal cancer proposed in 2015 by Guinney et al. Characterized by mesenchymal features, activation of TGF-β signaling and epithelial-mesenchymal transition (EMT), resistance to chemotherapy and immunotherapy, and the worst prognosis.

Note 2) EMT (Epithelial-Mesenchymal Transition): A biological process where epithelial cells acquire migratory and invasive mesenchymal traits, playing a key role in cancer invasion, metastasis, and treatment resistance.

Note 3) TGF-β Signaling Pathway: A critical signaling pathway involved in cell proliferation, differentiation, immune suppression, and stromal formation. It is hyperactivated in CMS4 colorectal cancer and contributes to therapy resistance via cancer-associated fibroblasts.

Note 4) T2-weighted imaging (T2WI) / Contrast-enhanced T1-weighted imaging (CE-T1WI): Standard MRI sequences. T2WI excels in assessing tissue heterogeneity, while CE-T1WI evaluates angiogenesis. Together, they provide complementary insights into the biological features of CMS4.

Note 5) SHAP (Shapley Additive Explanations): A method to quantitatively assess the contribution of each feature to a machine learning model’s prediction, enhancing model interpretability.

Note 6) Neoadjuvant chemoradiotherapy (nCRT): A treatment combining chemotherapy and radiation therapy administered before surgery. It is a standard approach, particularly for rectal cancer.

Paper Information

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

Authors: Zonglin Liu, Wenchao Gu, Liheng Liu, Shiman Wu, Zhenwei Yao, Yiqun Sun, Dan Huang, Tong Tong

Journal: Radiology

DOI: 10.1148/radiol.251719