Preliminary diagnosis of Parkinsons disease based on convolutional neural network

Abstract

The mainstream examination for Parkinsons disease is still to determine the Unified Parkinsons Disease Rating Scale (UPDRS), it can be inaccurate due to doctors or patients subjectivity. But recent studies have shown that patients of Parkinsons disease will represent a certain extent of dysgraphia in the early stage. Based on this feature, we proposed a method to realize the preliminary diagnosis for Parkinsons disease based on patients hand-drawings. In this paper, after the images that we used are pre-processed based on Threshold Segmentation, we set a 4-layer network for Convolutional Neural Networks (CNNs). First, we design a convolutional layer to learn the local features of hand-drawing, then the image is passed to then Maxpooling to go through the maximum pooling operation to preserve the contour features and to remove extraneous information. We set up two fully connected layers to capture more nonlinear relationships between images and labels. In the last, an accuracy calculation formula is adopted to diagnose Parkinsons disease. Overall, this diagnosis scheme that only requires patients hand-drawing can be completely automatic and more convenient than the traditional examination, the accuracy of result can be further improved if more details in the hand-drawing can be gathered.