MetaViT: Metabolism-Aware Vision Transformer for Differential Diagnosis of Parkinsonism with ¹⁸F-FDG PET

Accurate and early differential diagnosis of parkinsonism (idiopathic Parkinson's disease, multiple system atrophy, and progressive supranuclear palsy) is crucial for informing prognosis and determining treatment strategies. Current automated differential diagnosis methods for F-18-fluorodeoxyglucose (F-18-FDG) positron emission tomography (PET) scans, such as convolutional neural networks (CNNs), often focus on local brain regions and do not explicitly model the complex metabolic interactions between distinct brain regions. These interactions, as reflected in FDG PET images, are keys for the differential diagnosis of parkinsonism. Vision transformer (ViT) models are promising in modeling such long-range dependencies, but they may overlook the local metabolic alternations and have not been widely adapted for 3D medical image classification due to data limitations. Therefore, we propose a novel metabolism-aware vision transformer (MetaViT), which uses self-attention and convolution to explicitly characterize both global and local metabolic interactions between interrelated brain regions. A masked image reconstruction task is introduced to guide the MetaViT model to focus on disease-related brain regions, addressing the scarcity of 3D medical imaging data and improving the trustworthiness and interpretability of the resulting model. The proposed framework is evaluated on a 3D FDG PET imaging dataset with 902 subjects, achieving a high accuracy of 97.7% in the differential diagnosis of parkinsonism and outperforming several state-of-the-art CNN and ViT-based approaches.