Imaging-based transformer model predicts early therapy response in advanced nasopharyngeal carcinoma: a dual-center study.

INTRODUCTION

Deep-learning methodologies for predicting early response in locally advanced nasopharyngeal carcinoma (LA-NPC) remain unvalidated, with techniques like 2.5-D imaging and Transformers underexplored.

MATERIALS AND

METHODS

MRI images from LA-NPC patients diagnosed between January 2020 and March 2024 at two centers were analyzed. Patients (n = 184) were split into training (n = 89), validation (n = 39), and test (n = 56) sets. Three segmentation models-SegResNet, Unet, and UnetR-automatically delineated regions of interest (ROIs). A 2.5D approach integrated adjacent tumor sections into a transfer learning framework, leading to three predictive models: Clinical, Transformer, and Combined. Performance was assessed using ROC curves, calibration curves, and decision curve analysis (DCA).

RESULTS

The Transformer model outperformed others, achieving AUCs of 0.968, 0.957, and 0.830 for the training, validation, and test sets, respectively. The Clinical model had lower AUCs (0.898, 0.759, 0.658). The Combined model, integrating clinical data, matched or exceeded Transformer performance, particularly in the test set (AUC = 0.874).

CONCLUSION

The Combined model, leveraging Transformer architecture and clinical factors, demonstrates strong efficacy in predicting early response in LA-NPC patients undergoing chemoradiotherapy, suggesting its potential for improved personalized treatment.

CRITICAL RELEVANCE STATEMENT: This study critically validates a novel 2.5-D radiomic-Transformer fusion model that improves early response prediction for locally advanced nasopharyngeal carcinoma, directly advancing personalized chemoradiotherapy planning in clinical radiology.

KEY POINTS: Early treatment response prediction in locally advanced nasopharyngeal carcinoma lacks validated deep learning models using 2.5D imaging and Transformers. Transformer-based model achieved superior predictive performance compared to clinical or combined models. Integrating clinical data with Transformer imaging analysis improves personalized chemoradiotherapy outcome prediction.