To develop and externally validate machine learning radiomic models in a two-center retrospective study based on multiparametric MRI, including apparent diffusion coefficient (ADC) mapping, for distinguishing benign from malignant sinonasal tumors. This retrospective study enrolled 497 patients with pathologically confirmed sinonasal tumors from two centers (Center 1, n = 318; Center 2, n = 179). Data from Center 1 were randomly divided into training (70%) and internal test (30%) sets, while Center 2 served as the external validation cohort. Tumor ROIs on T1-weighted, T2-weighted, contrast-enhanced T1-weighted (CE-T1WI), and ADC images were manually segmented. A total of 1,319 radiomic features were extracted after min–max scaling, followed by Pearson correlation filtering, ANOVA/Kruskal–Wallis testing, and recursive feature elimination to select 1–20 key features. Ten classifiers were evaluated using nested cross-validation, and the final model was locked before testing in the internal and external cohorts. Multisequence models consistently outperformed single-sequence inputs. For single sequences, the best models and AUCs (training/internal test/external validation) were: T1WI (autoencoder, 2 features: 0.787/0.769/0.754), T2WI (logistic regression, 8 features: 0.825/0.818/0.740), CE-T1WI (naïve Bayes, 1 feature: 0.811/0.765/0.782), and ADC (Gaussian process, 1 feature: 0.897/0.884/0.875). Performance improved with feature fusion. The four-sequence model (Gaussian process, 3 features) achieved the highest AUCs of 0.932, 0.892, and 0.879. Among two- and three-sequence combinations, T2WI + ADC (0.905/0.871/0.883) and T2WI + CE-T1WI + ADC (0.910/0.890/0.877) were the top performers. Multiparametric MRI radiomics, especially combined-sequence models, enables accurate differentiation between benign and malignant sinonasal tumors and shows robust external generalizability to support clinical decision-making.
Sui et al. (Thu,) studied this question.