Noninvasive Characterization of Tumor Heterogeneity in Locally Advanced Nasopharyngeal Carcinoma: From Progression‐Free Survival Prediction to Biological Correlates

BACKGROUND: Locally advanced nasopharyngeal carcinoma (LA-NPC) shows substantial variability in progression-free survival (PFS) despite radiotherapy-based treatment. Clinical factors and conventional radiomics may not adequately characterize the spatial organization of intratumoral heterogeneity. Whether topology-aware MRI measures are associated with PFS and transcriptomic features remains unclear. PURPOSE: To evaluate an MRI-derived intratumoral heterogeneity (ITH) framework for PFS risk stratification and explore its transcriptomic correlates in LA-NPC. STUDY TYPE: Retrospective prognostic modeling study with internal and independent external validation and a separate prospective RNA-sequencing cohort. POPULATION: The retrospective cohort included 1001 patients (747 men and 254 women): training, n = 562; internal validation, n = 241; and external validation, n = 198. The prospective RNA-sequencing cohort included 53 patients (38 men and 15 women). FIELD STRENGTH/SEQUENCE: Pretreatment 3.0-T axial contrast-enhanced T1-weighted fast/turbo spin-echo MRI. ASSESSMENT: Primary tumors were manually segmented and reviewed by two senior radiologists. 2D and 3D ITH scores and conventional radiomics features were evaluated using survival models. Transcriptomic correlates were assessed using differential expression, co-expression network, and gene set enrichment analyses. STATISTICAL TESTS: Model discrimination was assessed using the concordance index (C-index) with 95% confidence intervals. Shapley additive explanations (SHAP) were used for model interpretation. Transcriptomic analyses used Benjamini-Hochberg false discovery rate (FDR) correction. RESULTS: The integrated heterogeneity-clinical DeepSurv model yielded C-indices of 0.727 (95% CI, 0.674-0.782) in the internal validation cohort and 0.714 (95% CI, 0.661-0.770) in the external validation cohort. The 3D ITH score had the largest mean absolute SHAP value (0.0554) among the model predictors. High 3D ITH scores were associated with enrichment of cytokine-cytokine receptor interaction (normalized enrichment score [NES] = 1.57; FDR = 0.00705) and viral protein interaction with cytokine and cytokine receptor signaling (NES = 1.77; FDR = 0.00264). DATA CONCLUSION: MRI-derived 3D ITH contributed to retrospective PFS risk stratification and showed exploratory cytokine-related transcriptomic correlates that require independent validation. EVIDENCE LEVEL: 3. TECHNICAL EFFICACY: Stage 2.