Parkinson's Disease Diagnosis Using Dual Attention Multi-Feature Network from Neuroimaging Data

Parkinson's disease (PD) is a progressive neurodegenerative disorder that requires early and accurate detection to improve patient outcomes. This study evaluates various deep learning and hybrid models for PD classification using structural MRI and DaTscan images independently. Pretrained architectures, including MobileNet, DenseNet121, AlexNet, ResNet50, and custom CNNs, were employed along with hybrid combinations such as MobileNet + DenseNet121 and CNN + ResNet50. Additionally, a novel Dual Attention Multi-Feature Network (DAMFN) block was integrated to enhance feature extraction by combining multi-scale self-attention, two-stage channel attention, and traditional convolution operations. The models were evaluated using accuracy, precision, recall, and F1-score, and Grad-CAM++ was applied to visualize discriminative regions in the images. Experimental results demonstrate that ResNet50 and hybrid models such as MobileNet + DenseNet121 achieve the highest classification performance on MRI images, while DenseNet121 and MobileNet also perform robustly on DaTscan data, with accuracy reaching up to 98% for MRI and 94.7% for DaTscan. The findings highlight the effectiveness of DAMFN-enhanced architectures for reliable PD prediction and provide insights into the spatial patterns relevant to disease diagnosis.