Advancing Diagnostic Imaging: AI, Radiomics, and Interpretability in Nasopharyngeal Carcinoma

The multimodal MRI radiomics model accurately predicts response to neoadjuvant chemotherapy in nasopharyngeal carcinoma and could change neoadjuvant-treatment selection in multidisciplinary care.

By integrating DCE‑MRI with standard clinical variables and moving from descriptive biomarkers to a predictive framework, the model departs from prior reliance on qualitative reads and isolated metrics. Radiologists and medical oncologists will now interpret case-level risk scores, shifting imaging from passive characterization to active treatment stratification and improving point-of-care decision support.

In a retrospective cohort of 370 patients (126 with DCE‑MRI), response to neoadjuvant chemotherapy was the primary endpoint. Across machine-learning and deep-learning approaches the model reported an AUC of 0.885, sensitivity 79.6%, and specificity 83.7%. That level of discrimination indicates the model can enrich selection of likely responders and reduce exposure of probable nonresponders to ineffective preoperative systemic therapy.

SHapley Additive exPlanations (SHAP) attributed prediction weights to imaging features, clarifying which radiomics variables drive individual-case outputs and supporting clinician trust. SHAP prioritized perfusion-heterogeneity measures from late DCE‑MRI enhancement phases, enabling targeted visual reassessment and hypothesis-driven tumor-board discussion; in practice, this interpretability helps reconcile algorithm output with clinical judgment during case review.

Adding clinical variables—stage, baseline laboratory indices, performance status—produced a more robust predictor than imaging alone, improving classification stability and calibration across subgroups. Integrated models reduced misclassification versus imaging-only models, underscoring the need for standardized capture of baseline clinical data so models receive consistent inputs for reliable patient-level predictions.

Limitations include the retrospective design, need for external prospective validation, and data-privacy and transferability constraints; each has paired mitigations. Multicenter prospective studies will test generalizability; federated-learning or secure anonymized-sharing frameworks can protect privacy while enabling cross-site training; and preplanned calibration plus scanner-harmonization protocols can reduce site effects. Prospective validation reports and pilots that embed the model into multidisciplinary workflows will determine whether successful validation and secure deployment shift preoperative treatment selection and clinical pathways for NPC management.