Radiomics of the midbrain on TCCD for identifying Parkinson's disease in substantia nigra hyperechogenicity-negative individuals.

This study evaluated the diagnostic utility of midbrain radiomic features from transcranial color Doppler (TCCD) imaging for identifying Parkinson's disease (PD) in individuals without substantia nigra hyperechogenicity (SN-). A total of 61 SN- PD patients and 61 age- and sex-matched SN- healthy controls were retrospectively enrolled after excluding subjects with visible hyperechogenicity or poor image quality. Midbrain regions of interest were manually segmented, and 464 radiomic features were extracted. Feature selection was performed using the Mann-Whitney U test with FDR correction and LASSO regression, followed by training of four machine learning models with five-fold cross-validation and independent testing. SHAP analysis was applied for model interpretation. All models performed well in the training set (e.g., XGBoost AUC = 0.96, SVM AUC = 0.91), but only the support vector machine (SVM) maintained stable performance on the test set, achieving an AUC of 0.79, accuracy of 0.74, and F1-score of 0.75. In contrast, XGBoost and random forest showed reduced performance, suggesting overfitting. Based on its consistent and balanced results, SVM was selected as the optimal classifier. SHAP analysis identified DependenceVariance, HighGrayLevelZoneEmphasis, ZoneVariance, and GrayLevelNonUniformity as the most influential features, reflecting heterogeneity, gray-level variability, and high-intensity zone prominence in the SN- PD group. These findings demonstrate that radiomics applied to midbrain TCCD images can identify PD in the absence of conventional ultrasound markers, providing an interpretable, noninvasive approach that may complement current diagnostic strategies.