Prediction of outcome in Parkinson's disease patients from DAT SPECT images using a convolutional neural network

Dopamine transporter (DAT) SPECT imaging is widely used for the diagnosis of Parkinson's disease (PD). Investigations on quantitative analysis of DAT images have been performed using radiomic features together with non-imaging features to predict a patient's outcome. The purpose of this study is taking the entire DAT image, without feature extraction, predicting a patient's motor function indicated by part III of the unified Parkinson's disease rating scale (UPDRS) using a convolutional neural network (CNN). We cast the motor function score prediction as a categorization problem that decides on whether a given patient will have a score below or above 30 in 4 years from baseline. This prediction practice was first conducted using only the baseline UPDRS_III score as the input, which resulted in an accuracy of 64.5%. The performance was then evaluated using the baseline score with the DAT image to assess whether the latter has added value for prediction. The resulting mean accuracy was significantly improved, reaching to 70.7%. This CNN-based motor function score predicting scheme demonstrates the contribution of DAT SPECT images, which shows their promise in improving progression tracking for patients with PD.