A Multiscale Hybrid Attention Networks Based on Multiview Images for the Diagnosis of Parkinson’s Disease

Parkinson’s disease (PD) is one of the common neurodegenerative diseases of the elderly. However, the modern healthcare lacks the apparatus to detect the early signs of the disease, with only selected experts being able to spot the onset. Therefore, the early detection of PD is particularly important. Convolutional neural networks, a deep learning technique that can automatically extract image features, have been widely used in the diagnosis of medical images. Due to the complexity of the organization in the brain, we proposed a multi-scale hybrid attention network (MSHANet) for automatic detection of healthy and Parkinson’s disease patients. MSHANet consisted of designed multi-scale convolutional blocks and introduced hybrid attention blocks, so it can capture complex features in brain images. Two datasets were created using the images in the publicly available Parkinson’s Progression Markers Initiative (PPMI) dataset, where the SV_3Dataset consisted of axial slices located in the substantia nigra region, and the MV_3Dataset adds mid-sagittal slices and striatal slices on the basis of SV_3Dataset. For these two datasets, we proposed two different classification strategies, namely Parallel Network Classification (PNC) and Multi-Slice Fusion Classification (MSFC), to improve the classification performance of PD. After cross-validation experiments, the best results for the model using the PNC strategy achieved are 90.59% of accuracy, 90.59% of precision, 90.61% of recall, 90.6% of F1 score, and 0.956 of AUC. By analyzing the above results, the striatal slice in MV_3Dataset provides higher accuracy than the other two slices. Both PNC and MSFC improved the classification effect of MSHANet on PD and HC, and the effect of PNC was better. The PNC strategy is used to test the performance of MSHANet on the test set. The best result is that the accuracy rate is 94.11%, the accuracy rate is 94.18%, the recall rate is 94.16, the F1 value is 94.17%, and the AUC is 0.9585. Our proposed method can help clinicians in accurately diagnosing the PD.