Data Driven Intelligent Diagnostics for Parkinson’s Disease

Parkinson's disease (PD) is a common chronic occult neurological degeneration disease which is most likely to occur among middle-aged and old-aged people. Neuroprotective methods can significantly slow the PD progression, therefore early diagnosis and the treatment of the PD is crucial. As the second most common neurodegenerative disorder, Parkinson's disease has its identification based on clinical diagnoses. Early diagnosis is especially important due to the lack of radical treatment. In all early diagnostic methods, results of imaging diagnosis based on positron emission tomography (PET) imaging has the most outstanding accuracy. Deep learning methods based on convolutional neural networks performed well in diagnosing different diseases. To address the demand of early diagnosis, this paper applies several preprocessing methods to image data such as gray level transformation, histogram equalization, improved wavelet soft-threshold denoising and image enhancement, and proposes a deep learning model based on U-Net architecture with deformable convolution kernels. After comparing the results of both methods, we found out that deformable U-Net outperformed the previously built improved VGG-Net. These results show that U-Net with deformable convolution structure has a good diagnostic capability and a better performance than the improved VGG-Net.