A Comparative Analysis of ViT-Based Embeddings and Machine Learning Classifiers for Parkinson's MRI Diagnosis

Abstract

In this study, a feature extraction approach based on Vision Transformer was evaluated for the binary classification of Parkinson's disease from T1-weighted MRI images. Embeddings obtained from the ViT-B/16, ViT-B/32, ViT-L/16, and ViT-L/32 architectures were classified using machine learning models such as Linear SVM, Logistic Regression (Elastic Net), Random Forest, KNN, Extra Trees, Gradient Boosting, AdaBoost, Naive Bayes, and XGBoost. The results show that the ViT-L/16 architecture offers the most discriminative representation and, when combined with the XGBoost model, achieves the highest performance with an accuracy of 0.9604, an F1 score of 0.9690, and an AUC of 0.9917. The findings reveal that Transformer-based visual feature extraction is effective in capturing subtle structural differences in medical images and, when combined with classical machine learning methods, delivers a powerful hybrid performance. © 2026 IEEE.