Leveraging machine learning framework and GANs for Parkinson disease detection

Parkinson's Disease (PD) is a progressive disorder that affects the nervous system, and the parts of the body controlled by nerves. Early-stage detection of PD using spiral and wave images can significantly improve patient outcomes. Current research has identified limitations in the classification of PD such as reduced dataset size. Processing a large and varied dataset and achieving high accuracy in PD detection can be a challenge. This research proposes a machine learning framework to improve the early detection of Parkinson's Disease by improving the accuracy. This research creates a novel dataset called GAN-PD Hybrid Dataset that combines 1632 original handwritten images and 1868 GAN-generated images for both Parkinson's and healthy subjects, used to train hybrid (ResNet50 and InceptionV3 with KNN) and standalone CNN models (ResNet50, InceptionV3). Data pre-processing and transfer learning techniques are applied to two pre-trained CNN models, namely ResNet50 and InceptionV3. Each of these models is evaluated both individually and in combination with KNN classifier. Results of these models are presented in this paper based on accuracy, sensitivity, specificity, F1 score, Cohen’s Kappa and precision. This research shows promise for InceptionV3 in aiding medical practitioners by detecting PD at an early stage.