A swin transformer and CNN fusion framework for accurate Parkinson disease classification in MRI

Parkinson's disease ranks as the second most prevalent neurological disorder after Alzheimer's disease. Convolutional neural networks (CNNs) have been extensively employed in Parkinson's disease (PD) detection using MR images. However, CNN models generally focus on local features while prone to capture global representations. On the other hand, the vision transformer (ViT) excels at capturing global features through its self-attention mechanism, but it compromises local feature representations. Additionally, the varying magnitude of MR data poses a challenge for ViT, potentially leading to the gradient vanishing problem. To address these limitations, this paper proposed a novel framework that combines the Swin-Transformer and CNN to capture both local and global features effectively. To mitigate the gradient vanishing issue in ViT, we used skipped connections and cosine attention mechanism in VIT that preserves the output distribution regardless of input magnitude variations. The proposed model comprises three primary blocks: Transformer-block, convolutional block, and dense-block. The input image is processed concurrently by the cosine transformer and convolutional block. Subsequently, the extracted features from both blocks are concatenated and fed to the dense block for decision-making. The proposed model achieved promesing results of 96%, 97%, 95%, and 95% in terms of accuracy, sensitivity, specificity, and area under the curve, respectively.