Objective To investigate alterations in diffusion tensor imaging (DTI)-derived fractional anisotropy (FA) of the substantia nigra and insula in patients with Parkinson’s disease (PD), and to explore the internal classification performance and feature contribution of machine learning models based on FA features in a single-center case-control cohort. Methods Thirty-five patients with PD and 28 healthy controls (HCs) were included. Group differences in bilateral substantia nigra FA, bilateral insular FA, and derived FA indices were compared. Linear regression analyses were performed after adjustment for age, sex, and years of education. Elastic net logistic regression, radial basis function kernel support vector machine, and random forest models were constructed using four original FA features. Model performance was evaluated using nested cross-validation, and permutation importance was used to assess ferture contribution. Results Mean FA values of the bilateral substantia nigra, bilateral insula, and the overall average FA were significantly lower in the PD group than in the HC group (all P<0.001), and these differences remained significant after adjustment for covariates. The ACU for discriminating PD from HC using mean bilateral substantia nigra FA was 0.919. In nested cross-validation, the radial basis function kernel support vector machine achieved an AUC of 1.000. Permutation importance indicated that FA values of the right and left substantia nigra contributed more substantially to model discrimination. Conclusion In this single-center case-control cohort,ROI-FA was reduced in both the substantia nigra and insula in patients with PD, with more robust between-group differences observed in substantia nigra FA.