This study investigates the potential of DINOv3 as a deep radiomics feature extractor for MRI-based differentiation of peripheral nerve sheath tumors (PNSTs) in neurofibromatosis type 1. We analyzed 3D T2- weighted fat-suppressed MRI scans from two patient cohorts (122 and 36 patients) and compared four feature extraction approaches: classical radiomics, the foundation model for cancer imaging biomarkers (FMCIB), Radio DINO, and DINOv3. In the first experiment, we assessed feature generalizability using k-nearest neighbor classifiers in leave-one-patient-out cross-validation to quantify how well features selected for one tumor differentiation task transfer to another. In the second experiment, we reproduced the prior PNST differentiation benchmark using random forest classifiers to evaluate the clinical relevance of deep radiomics features. DINOv3 demonstrated robust feature generalization and achievedAUCof 0.76–0.96 across tasks in the PNST differentiation benchmark. It consistently outperformed classical radiomics (AUC 0.60–0.94) and FMCIB (AUC 0.60–0.84), and performed on par with Radio DINO (AUC 0.82–0.94). These findings indicate that DINOv3 embeddings provide generalizable and clinically discriminative deep radiomics features for PNST differentiation in NF1.