Development and validation of an interpretable MRI-based multimodal fusion model for predicting lymph node metastasis after neoadjuvant chemoradiotherapy in locally advanced rectal cancer: a multicenter study.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiology, The First Peoples Hospital of Foshan (Foshan Hospital Affiliated to Southern University of Science and Technology), School of Medicine, Southern University of Science and Technology, Foshan, China.
- Department of MRI, Maoming Peoples Hospital, Maoming, China.
- Department of Radiology, Foshan Hospital of Traditional Chinese Medicine, Foshan, China.
- The Eighth Clinical Medical College of Guangzhou University of Chinese Medicine, Foshan, China.
- Department of Radiology, Yuebei Peoples Hospital, Shaoguan, China.
- Department of information system, The First Peoples Hospital of Foshan (Foshan Hospital Affiliated to Southern University of Science and Technology), School of Medicine, Southern University of Science and Technology, Foshan, China.
- Department of MRI, Maoming Peoples Hospital, Maoming, China. [email protected].
- Department of Radiology, The First Peoples Hospital of Foshan (Foshan Hospital Affiliated to Southern University of Science and Technology), School of Medicine, Southern University of Science and Technology, Foshan, China. [email protected].
Abstract
To develop an interpretable multimodal model that integrates pre-treatment Magnetic resonance imaging (MRI)-based deep learning radiomics (DLR) features with Node-RADS scores to predict Lymph node metastasis (LNM) in locally advanced rectal cancer (LARC) patients after neoadjuvant chemoradiotherapy (nCRT). In total, 190 LARC patients from three independent centers were retrospectively included. Deep learning and radiomics features were derived from T2-weighted and diffusion-weighted imaging sequences before treatment. Lymph nodes were scored by radiologists using the Node-RADS system. The DLR Node-RADS (DLRN) combined model was constructed using the logistic regression classifier, and receiver operating characteristic analysis was employed to evaluate model performance. Furthermore, feature contributions were clarified using the SHapley Additive exPlanations (SHAP) analysis method. The DLRN joint model has superior performance in predicting LNM, and it outperforms the individual DLR model and the Node-RADS model in both the training and two external validation sets. For training set, the area under the curve (AUC) were 0.941 for the DLRN model, 0.864 for the DLR model, and 0.852 for the Node-RADS model. In external validation set 1, the AUCs were 0.923, 0.810, and 0.765, respectively; in external validation set 2, they were 0.843, 0.817, and 0.801. SHAP analysis determined that Node-RADS and DL features are among the most critical predictive features. The DLRN model based on multiparameter MRI has demonstrated a promising predictive ability for evaluating the LNM of LARC patients after nCRT. The novel interpretable framework enhances the clinical credibility of the model in guiding personalized decisions for rectal cancer.