Liver metastasis remains the most common and life-threatening progression in rectal cancer, substantially lowering patient survival. Although rectal MRI and abdominal CT are routinely used for diagnosis and staging, their limitations in early detection of micrometastases hinder effective treatment planning. In response, a novel multi-modal radiomics model combining rectal MRI and pre-metastatic liver CT has been developed to improve early prediction of liver metastasis in rectal cancer patients. This explainable model, validated across multiple centres, offers both predictive accuracy and prognostic value, supporting personalised care strategies and better-informed clinical decisions.
Model Design and Development
The study retrospectively enrolled 431 patients with pathologically confirmed rectal cancer from two medical centres. Patients underwent baseline rectal MRI and contrast-enhanced liver CT within two weeks prior to treatment initiation. Radiomics features were extracted from T2-weighted and diffusion-weighted MRI images of the primary tumour, as well as from the entire liver parenchyma on CT images. A rigorous feature selection strategy was applied, starting with ANOVA F-value ranking to identify the top 20% of radiomic features, followed by recursive feature elimination using a support vector machine (SVM) estimator. This process aimed to minimise redundancy while preserving meaningful data.
To address class imbalance between liver metastasis and non-metastasis groups, the Synthetic Minority Oversampling Technique (SMOTE) was employed. Separate models were constructed using MRI-only features, CT-only features and a fused feature set combining both modalities. A clinical model was also developed based on selected variables such as age, carcinoembryonic antigen (CEA), CA19_9, tumour distance from the anus, tumour length, N stage, mesorectal fascia involvement, tumour deposits and extramural venous invasion. These were selected via LASSO regression. All models were evaluated using receiver operating characteristic (ROC) curve analysis, with area under the curve (AUC) and other performance metrics assessed in training, internal validation and external validation cohorts.
Predictive Performance and Model Explainability
The fusion radiomics model achieved strong performance in predicting liver metastasis. In the training cohort, it yielded an AUC of 0.85, which slightly decreased to 0.75 and 0.73 in internal and external validation cohorts, respectively. When combined with clinical variables, the model demonstrated even higher accuracy, reaching an AUC of 0.91 in the training cohort and maintaining robust performance in validations. Comparatively, the clinical model alone had limited predictive capability, especially in the external validation cohort, where its AUC dropped to 0.55. In contrast, the combined model maintained high sensitivity, specificity and net clinical benefit, as confirmed by decision curve analysis.
Must Read: Comparing MRI Protocols for Colorectal Liver Metastases Detection
To enhance interpretability, the SHAP (Shapley Additive Explanations) method was applied to visualise the impact of each feature on the model’s output. Key radiomic features such as glcm_Imc2 and firstorder_Kurtosis were identified as influential in distinguishing between metastasis and non-metastasis cases. For instance, lower values of glcm_Imc2 were associated with a higher probability of liver metastasis, indicating a potential link to microenvironmental changes in liver tissue. SHAP force plots and waterfall plots further illustrated how individual features contributed to prediction outcomes in representative patient cases. A nomogram was also constructed to visualise the combined model and guide clinical interpretation.
Prognostic Value and Clinical Implications
Beyond predicting metastasis, the model’s RadLM score effectively stratified patients into high- and low-risk groups for liver metastasis-free survival (LMFS). Kaplan–Meier survival analysis showed significantly shorter LMFS in the high-risk group across both training and internal validation cohorts. Multivariate Cox regression confirmed that the RadLM score was an independent prognostic factor, with a hazard ratio of 5.50. Subgroup analysis further supported the model’s prognostic value, independent of post-operative chemotherapy status.
The integration of both primary tumour and liver features provides a comprehensive understanding of metastatic potential. Unlike previous studies that focused solely on the primary tumour or used limited liver data, this study leveraged automated segmentation of the entire liver to capture radiomic signatures potentially indicative of a pre-metastatic niche. The presence of liver tissue heterogeneity prior to visible metastasis supports the theory that microenvironmental changes precede radiological detection, making this model particularly useful for early intervention planning.
The multi-modal, explainable radiomics model presented in this study represents a significant advance in early prediction of liver metastasis in rectal cancer. By integrating rectal MRI and liver CT features, the model demonstrates high predictive accuracy and adds prognostic value through the RadLM score. SHAP analysis and nomogram visualisation support interpretability and clinical usability. These findings suggest that early identification of high-risk patients is feasible, allowing for tailored treatment strategies and more intensive follow-up. Despite limitations such as imaging variability and retrospective design, the study underscores the potential of radiomics in precision oncology and encourages further exploration into multi-organ imaging biomarkers for metastasis prediction.
Source: Insights into Imaging
Image Credit: iStock
