Combining Convolutional Neural Network (CNN) and Grad-CAM for Parkinson’s Disease Prediction and Visual Explanation

Parkinson's disease is one of the types of neurological diseases that is caused by the destruction of brain cells that produce dopamine. Early detection of Parkinson's disease is an important factor in slowing the progression of the disease. In this study, a Convolutional Neural Network (CNN) namely ConvNet, is used to discriminate Parkinson's patients based on Single Photon Emission Computed Tomography (SPECT) images acquired from the PPMI database. Since the dataset is limited, after a pre-processing stage, two data augmentation techniques are used. Finally, the Grad-CAM technique is used to obtain visual interpretation from the predictions of the proposed CNN. To evaluate the proposed method, different measures such as accuracy, sensitivity (recall) and f1-score are used. Simulation results according to the measures shows that when the classic data augmentation method is used accuracy is increased to 98.50% and more efficient classification is performed.