COMPUTATIONAL VISION IN THE SERVICE OF DIAGNOSING PARKINSON'S DISEASE

Abstract

This chapter investigates the application of deep learning models for the diagnosis of Parkinson's disease (PD) from motor test images. The study explores transfer learning techniques in convolutional neural networks and compares their performance with a classic random forest model. Three pre-trained convolutional neural network architectures were explored, fine-tuned for binary classification indicating the presence or absence of PD. For comparison, a random forest model with feature extraction using a histogram of oriented gradients was implemented. The dataset contains 204 images with wave and spiral test images, which were divided into 70% for training and 30% for testing. The evaluation metrics adopted include accuracy, sensitivity, and specificity. For wave-type drawings, MobileNetV2 achieved the highest accuracy (89.9%), while ResNet50V2 achieved the best sensitivity (93.3%). InceptionV3 excelled in specificity (100%). For spiral drawings, ResNet50V2 showed the best overall performance (86.6% accuracy). The random forest achieved a maximum accuracy of 80.6%, lower than the neural networks. The results demonstrate that deep learning models with transfer learning are effective in aided diagnosis of PD. The choice of model may vary depending on clinical priority. The MobileNetV2 model maximizes accuracy, while ResNet50V2 maximizes sensitivity, and InceptionV3 minimizes false positives. The developed application illustrates the potential for integrating these techniques into clinical support tools.

DOI: https://doi.org/10.56238/edimpacto2025.091-008