Parkinson Hastalığının Beyin MRI Görüntülerinden Tespitinde CNN ve Transformer Tabanlı Modellerin Karşılaştırılması

Altuncu M. A., Kaplan K.

Fırat Üniversitesi Mühendislik Bilimleri Dergisi, cilt.38, sa.2, ss.645-654, 2026 (TRDizin)

Özet

Parkinson’s disease is an important neurological disorder whose diagnosis can be challenging, especially in the early stages, due to the lack of pronounced symptoms. Therefore, interest in computer-assisted methods for early and accurate disease detection has been increasing. In this study, four deep learning models were comparatively evaluated for the detection of Parkinson’s disease using brain magnetic resonance images (MRI). Specifically, the Baseline CNN, EfficientViT, DeiT, and MobileViT architectures were evaluated using the same dataset and common training settings. Models were evaluated using accuracy, weighted F1 score, sensitivity, specificity, number of parameters, training time, inference time, and FPS performance metrics. On the fixed test set, the MobileViT model achieved a value of 1.0000 in accuracy, weighted F1, sensitivity, and specificity metrics, while the EfficientViT model provided the best speed-performance balance with 0.992 accuracy, 3.35 ms/image inference time, and 298.21 FPS. The performance of the MobileViT model across different data splits was also examined using five-fold stratified cross-validation, yielding a result of 1.0000 ± 0.0000 in key performance metrics. The results indicate that MobileViT and EfficientViT show promising performance for MRI-based Parkinson’s disease classification. However, validation on larger, patient-level, multicenter datasets and further clinical evaluation are required before these models can be considered for use in clinical decision-support systems.