3D SPECT-based machine learning approach to early Parkinson’s diagnosis

Purpose:: Dopamine transporter binding measurements in the striatum play a critical role in identifying dopaminergic deficits in Parkinson’s disease (PD), particularly during early stages. Using 3D SPECT imaging data from the PPMI database (928 subjects), we developed an automated feature extraction pipeline that captures subtle morphological changes characteristic of early neurodegeneration. Approach:: Our methodology involved a sub-volume from each 3D SPECT scan and applied several image processing techniques to each slice. Unlike approaches relying on external datasets or derived features, we focused exclusively on direct geometric and shape features. We extracted 30 novel discriminative features, including volumetric measures, global thresholds, circularity, compactness, and curvature, with rigorous statistical validation using parametric tests (t-tests, ANOVA). The top 19 features, prioritized through MRMR to eliminate redundancy, were used to train and evaluate 34 machine learning classifiers. Results:: For binary classification (Healthy Control (HC) vs. early PD), our Medium Gaussian SVM achieved excellent testing performance (98.56% accuracy, 98.81% precision), significantly higher than recent state-of-the-art approaches. Meanwhile, for three-class classification (HC, PD stage 1, and PD stage 2), the Subspace Discriminant Ensemble achieved the highest accuracy of 76.26%, solving a critical real-world scenario. Conclusions:: This stage-aware approach advances the field by enabling automatic bilateral analysis of the striatum and accurate classification of early-stage PD, potentially supporting earlier clinical intervention. It also addresses a critical gap in the literature where most methods focus solely on binary classification (HC vs. PD) without capturing disease progression.