Radiomics‑based differentiation of renal tumor subtypes using [89 Zr]Zr‑Girentuximab PET and MRI

Clear-cell renal cell carcinoma (ccRCC) is the most common and aggressive subtype of renal malignancy. Accurate, non-invasive differentiation between ccRCC and non-clear-cell RCC (non-ccRCC) remains a clinical challenge. This study evaluates the utility of radiomic features extracted from [89Zr]Zr-girentuximab PET and T1-weighted MRI for subtype classification using machine learning approaches. A cohort of 17 patients (9 ccRCC, 8 non-ccRCC) underwent PET/MRI imaging. Tumor lesions were manually segmented, and 1,874 radiomic features were extracted from each modality using PyRadiomics. Z-score normalization was applied, and four supervised classifiers including random forest (RF), support vector machine with radial basis function kernel (svmRadial), logistic regression (LogReg), and generalized linear model via elastic net regularization (GLMNet), which were trained and evaluated using leave-one-out cross-validation. Models trained on PET-derived features achieved the highest performance across all metrics. GLMNet attained the best classification results with an area under the curve (AUC) of 0.972, accuracy of 0.882, and perfect specificity (1.0). MRI-based models yielded lower and more variable performance, with reduced AUC and specificity. Combining PET and MRI features did not improve results beyond those of PET alone, suggesting limited complementary value from MRI in this setting. Visual analysis of radiomic heatmaps further supported the superior discriminative power of PET features. Among classifiers, GLMNet and LogReg showed consistent robustness, while svmRadial underperformed. These findings highlight the value of functional PET imaging over structural MRI for ccRCC subtype classification. Future studies should explore the integration of foundation models trained on large multimodal datasets to improve generalizability in small-sample radiomics research.