Comprehensive Summary
This study aimed to construct and validate an interpretable radiomics model based on pelvic magnetic resonance imaging (MRI) for the preoperative prediction of lateral lymph node metastasis (LLNM) in patients with lower rectal cancer (LRC). A total of 253 patients with pathologically confirmed LRC were retrospectively enrolled, split into a training cohort (n=177) and testing cohort (n=76). Radiomic features were extracted from axial T2-weighted MRI images, and a machine learning pipeline was developed using LASSO regression, minimum redundancy maximum relevance (mRMR), and XGBoost for feature selection and classification. Six radiomic features were ultimately selected for the model. The final XGBoost-based model achieved AUCs of 0.845 in the training set and 0.812 in the testing set. Calibration and decision curve analyses confirmed its clinical utility. An interpretable SHAP (SHapley Additive exPlanations) framework was integrated to visualize individual prediction contributions of each feature, enhancing transparency for clinical decision-making.
Outcomes and Implications
This study presents a clinically relevant, non-invasive tool with the potential to reshape preoperative planning in lower rectal cancer. The radiomics model achieved strong predictive power with an AUC of 0.845 in the training cohort and 0.812 in the testing cohort, while maintaining specificities of 81.2% and 75.0% and accuracies of 80.2% and 76.3%, respectively. The model’s integration of SHAP analysis enabled transparency in feature influence, enhancing its interpretability in clinical settings. These findings support its potential in optimizing decisions regarding lateral lymph node dissection (LLND), which is associated with high surgical risk yet remains controversial. Importantly, patients with LLNM had significantly worse outcomes, underscoring the value of early identification.