MRI–Based Multilevel Radiomics and Transformer Features for Predicting Radiation–Induced Carotid Artery Injury after Nasopharyngeal Carcinoma Radiotherapy: A Multicenter Study
Abstract
Purpose
To develop and validate an MRI–based fusion model (Rad–SRad–SwinT) integrating conventional radiomics (Rad), subregional radiomics (SRad), and Transformer–derived deep learning features (Swin Transformer, SwinT) to predict post–radiotherapy radiation–induced carotid artery injury (RICAI) in nasopharyngeal carcinoma (NPC).
Methods
In this multicenter retrospective study, 500 NPC patients from four hospitals were allocated to training (n=274), internal testing (n=118), and external testing cohorts (n=108). Rad features were extracted from MRI–defined carotid artery regions of interest, SRad features from K-means–derived subregions, and deep features from a SwinT backbone. Single-source and fusion models were developed. Discrimination (AUC), classification (ACC/SEN/SPE), calibration (Brier score and calibration curves), reclassification (NRI/IDI), and interpretability (SHAP) were assessed.
Results
RICAI was observed in 48.5%, 48.3%, and 54.6% of the training, internal testing, and external testing cohorts, respectively. Among single–source models, SwinT and SRad showed comparable performance, with Rad slightly inferior; all outperformed the clinical model. The fused Rad–SRad–SwinT achieved the best performance, with AUCs of 0.814 (95% CI: 0.737–0.891) in internal testing and 0.871 (95% CI: 0.794–0.932) in external testing, alongside favorable classification in external testing (ACC 0.815, SEN 0.763, SPE 0.878) and good calibration (Brier score 0.148). NRI/IDI analyses indicated significantly improved reclassification versus single-source models. SHAP analyses demonstrated that SwinT-derived features contributed most to model decisions, followed by SRad and Rad, supporting complementary gains from deep semantic representation and subregional heterogeneity quantification.
Conclusion
Integrating multilevel radiomics with Transformer–derived deep learning features enhances prediction of RICAI after NPC radiotherapy and shows promise as a noninvasive risk–stratification tool.