Deep learning-based prediction of post-radiation nasopharyngeal necrosis using adaptive multimodal fusion of CT, MRI, and 3D dose

This study aimed to develop and internally validate a deep-learning model for the risk stratification of post-radiation nasopharyngeal necrosis within five years of intensity-modulated radiotherapy for nasopharyngeal carcinoma, a rare but life-threatening complication. We retrospectively analysed 170 patients with nasopharyngeal carcinoma who received primary intensity-modulated radiotherapy; 62 developed post-radiation nasopharyngeal necrosis. Pre-treatment inputs comprised planning computed tomography, contrast-enhanced T1-weighted magnetic resonance imaging, T2-weighted magnetic resonance imaging, and the three-dimensional dose distribution. The framework uses four modality-specific three-dimensional encoders, a cross-modal Transformer, and adaptive mixture-of-modality-experts fusion. Performance was estimated by leakage-controlled five-fold stratified cross-validation, with model selection performed on an internal validation split of each training fold. We report 95% bootstrap confidence intervals and use the DeLong test for pairwise model comparisons. The framework achieved an area under the receiver operating characteristic curve of 0.907 (95% confidence interval 0.861–0.948), accuracy 0.876 (0.824–0.924), sensitivity 0.888 (0.809–0.954), specificity 0.872 (0.808–0.927), positive predictive value 0.808 (0.740–0.888), negative predictive value 0.933 (0.893–0.973), and Brier score 0.140 (0.119–0.161). Adding both magnetic resonance imaging sequences to dose and computed tomography increased the area under the curve from 0.695 to 0.907. Adaptive expert fusion outperformed a matched concatenation baseline with an area under the curve of 0.798 (DeLong \(p<0.001\) ). It also outperformed four of five re-implemented baselines ( \(p\leq0.009\) ); its increase of 0.036 over the strongest baseline did not reach conventional significance ( \(p=0.089\) ). Apparent decision curve analysis showed positive net benefit over a wide range of threshold probabilities in the enriched study sample. The proposed framework provides discriminative patient-level estimates of post-radiation nasopharyngeal necrosis risk in internal five-fold cross-validation, with the lowest Brier score among the compared models but with apparent miscalibration of absolute probabilities on this enriched sample. External prospective validation, probability recalibration, and evidence linking risk-guided surveillance to improved clinical outcomes are required to establish clinical utility.