Parkinson’s disease detection using different cnn architectures with effective feature selection approach

Abstract

The most prevalent neurological condition affecting the central nervous system is Parkinson's disease. It has been claimed that PD sufferers' handwriting deteriorates. In this sector different machine learning algorithms and techniques are applied however these have some limitations. Because of model bias generated by data inequity, these artificial intelligence architectures perform well on the vast majority however poorly on the minority class. To address this issue, we propose a model where the training process is balanced using a random under sampling strategy. The re-sampling techniques are not used on the entire dataset before cross validation, rather solely on the training phase at every cross-validation iteration. In this purpose, we use the HandPD dataset, which is divided into two sections: Spiral data and Meander data. In this study, we use CNN to extract features from handwritten dynamics photos, which collect a range of data while analyzing the subject. The HandPD dataset is divided into two portions, 25% used for testing and 75% for training. with 128*128 and 64*64 images utilized in both cases. For comparison and to discover the optimum architecture, we propose CNN Architecture 1(CA1) and CNN Architecture 2(CA2). To serve as a point of reference, we run a second experiment on the original data. Despite the fact that any machine learning methodology can be used, we pick the OPF (Optimum-Path Forest) classifier because it is quick and parameter less. We calculate the overall accuracy and average control and Parkinson Disease patient accuracies across the entire test set for each meander dataset and spiral dataset separately. CA1 performs better in terms of total accuracy averaged for the test set when using the meander dataset, with accuracy of 87.24% and 85.10% for 128*128 and 64*64 images, respectively. For average Performance of Patients with Parkinson disease throughout test set using the meander dataset, we find that OPF performs better, with 93.66% and 91.66% accuracy for 128*128 and 64*64 images, respectively. Average overall accuracy across all test sets in the case of the Spiral dataset, we discover that OPF performs better, with an accuracy of 76.82% for 128*128 images, and CA1 performs better, with an accuracy of 81.19% for 64*64 images. Accuracies over the test set for average PD patients in the case of the Spiral dataset, we discover that CA2 performs better, with an accuracy of 90.89 percent for 128*128 images, while CA1 performs better, with an accuracy of 87.68 percent for 64*64 images. Experiments show that different CNN architectures are better in different scenarios. However, in the great majority of situations, CA1 performs better.