A Deep Learning-CNN Based System for Medical Diagnosis: An Application on Parkinson's Disease Handwriting Drawings

Abstract

Parkinson's disease (PD) is a degenerative disease that affects the motor system, which may cause slowness of the speech and the movements, and the anomaly of writing abilities due to tremor. PD diagnosis by Deep Learning approach has become an important worldwide medical issue through the last years. It is obvious that these patients due to their physical conditions are not suitable for every kind of PD diagnosis test. One of the non-invasive PD identification methods is the handwriting test, which is utilized in hospitals since many years ago. In this work we propose Convolutional Neural Network (CNN) based Deep Learning system to learn features from Handwriting drawing spirals which are drawn by People with Parkinson; also, we evaluated the performance of our deep learning model by K-Fold cross validation and Leave-one-out cross validation (LOOCV) techniques. Moreover, we introduce a dataset with a novel test which is called Dynamic Spiral Test (DST) along with traditional Static Spiral Test (SST) for PD recognition. We used both dynamic features and visual attributes of spirals. The proposed approach was reached to 88% accuracy value. The analysis of handwritten drawing tests proves that it is useful to combine SST and DST tests for automatic PD identification.