Multi-parameter magnetic resonance imaging (MRI) deep learning radiomics predicts complete response after induction immunochemotherapy in locally advanced nasopharyngeal carcinoma.

Background: Concurrent chemoradiotherapy (CCRT) following induction chemotherapy (IC) remains the conventional treatment regimen for patients with locally advanced nasopharyngeal carcinoma (LANPC). However, the complete response (CR) rate after IC is limited, and there is considerable heterogeneity in patient responses. The introduction of immunotherapy has shown potential to enhance treatment efficacy; nevertheless, there is a lack of reliable biological markers that can effectively predict whether induction immunochemotherapy will result in CR. Multiparametric magnetic resonance imaging (MRI) offers a non-invasive method to provide comprehensive information regarding tumor structure and function. This study aims to develop and validate a multimodal fusion model that integrates traditional MRI-based radiomics features with deep learning radiomics features to predict the early achievement of CR in patients with LANPC undergoing induction immunochemotherapy, thereby establishing a foundation for personalized treatment decisions. Methods: We conducted a retrospective analysis of clinical and imaging data from 230 biopsy-confirmed LANPC patients who underwent induction immunochemotherapy at Guangxi Medical University Cancer Center between January 2021 and December 2024. The patients were randomly allocated into training (n=184) and testing (n=46) cohorts. Regions of interest (ROIs) for the lesions were delineated across multiple sequences, including T1-weighted imaging (T1), T2-weighted imaging (T2), and contrast-enhanced T1-weighted imaging (CE-T1). Traditional and deep learning radiomics features were extracted, followed by feature selection to identify the most discriminative features. Utilizing machine learning algorithms, we developed four types of models: clinical, traditional radiomics, deep learning radiomics, and multimodal fusion models. Model performance was evaluated through receiver operating characteristic (ROC) curve analysis, area under the curve (AUC), and decision curve analysis (DCA). Results: The multimodal fusion model exhibited superior predictive performance in the testing cohort [AUC =0.844, 95% confidence interval (CI): 0.695-0.992], significantly outperforming both the traditional radiomics fusion model (AUC =0.721; 95% CI: 0.540-0.901) and the deep learning radiomics fusion model (AUC =0.725; 95% CI: 0.566-0.885). Conclusions: The multimodal fusion model effectively predicts early CR in LANPC patients following induction immunochemotherapy, demonstrating significant potential for clinical application.