A Multimodal MRI-Based Model for Colorectal Liver Metastasis Prediction: Integrating Radiomics, Deep Learning, and...

Area of Science:

• Oncology
• Radiology
• Medical Imaging
• Machine Learning

Background:

• Predicting colorectal cancer liver metastasis (CRLM) is crucial for patient prognosis.
• Multiparametric MRI offers potential for noninvasive CRLM assessment.

Purpose of the Study:

• To develop and validate an interpretable multimodal machine learning framework using multiparametric MRI for CRLM prediction.
• To enhance model interpretability using SHapley Additive exPlanations (SHAP) and deep learning visualization.

Main Methods:

• A multicenter retrospective study of 463 colorectal cancer patients.
• Extraction of radiomics features from T2WI and DWI, and deep learning features using ResNet101.
• Development of clinical, radiomics, deep learning, and combined models using LASSO logistic regression.

Main Results:

• The combined model achieved high AUCs (0.889 training, 0.838 internal, 0.822 external validation), outperforming single modalities.
• SHAP analysis identified T2WI texture and CA19-9 as key predictors; Grad-CAM confirmed deep learning focus on tumor regions.
• Decision curve analysis indicated enhanced clinical net benefit for the integrated model.

Conclusions:

• A robust, interpretable multiparametric MRI model was developed for noninvasive CRLM prediction.
• Integration of radiomics and deep learning features with SHAP/Grad-CAM enhances predictive performance and clinical relevance.
• The model shows potential as a decision-support tool for individualized colorectal cancer management.