Development and validation of an interpretable deep learning radiomics model using MRI to predict lymph node metastasis in rectal cancer.

Motivation: Lymph node (LN) metastasis is essential for guiding treatment and predicting outcomes in rectal cancer. Goal(s): This study develops and validates an interpretable DLR model using multiparametric MRI to predict preoperative lymph node metastasis in rectal cancer. Approach: This retrospective study included 286 rectal cancer patients for training and 66 for testing. We extracted radiomics features from MRI scans, generated deep learning features, assessed interpretability using SHAP, and evaluated survival prediction with Kaplan-Meier curves. Results: The DLR model demonstrated strong predictive accuracy, with AUCs of 0.878 and 0.752, significantly enhancing radiologists' performance and effectively stratifying patients by disease-free survival in rectal cancer. Impact: This DLR model's accuracy and interpretability support improved diagnostic confidence in rectal cancer, aiding clinicians in decision-making. By bridging advanced imaging and clinical needs, this tool opens new possibilities for preoperative assessments and personalized oncology care.