Early Detection of Parkinson's Disease Based on Diffusion Tensor Imaging and Deep Learning

Parkinson's disease (PD) is the most common movement disorder and the second neurodegenerative disease. Before the appearance of symptoms, the biochemical and pathological changes start in the brain, which are not often easily appreciable in routine magnetic resonance imaging (MRI). Several studies have performed statistical analysis on fractional anisotropy (FA) and mean diffusivity (MD) maps extracted from diffusion tensor imaging (DTI) data to identify the differences between PD and healthy control (HC) groups. Their results demonstrated a significant reduction in fractional anisotropy (FA) and increased mean diffusivity (MD) in many brain regions of the PD group compared to the healthy control. Though some studies have shown Parkinson's segregation from healthy subjects based on a statistical analysis of FA and MD maps extracted from DTI data, the current diagnostic procedures are expertise-demanding and thus necessitate a classification of PD based on artificial intelligence (AI). In this study, we aim to investigate different deep learning solutions (using a 3D ResNet-18 neural network) trained with different frameworks. To this end, six different training scenarios were investigated through feeding the network with different input data extracted from the DTI to determine the best performing framework for early PD detection. The six model training frameworks include (1) single-input for only DTI imag es, (2) single-input for only FA maps, (3) single-input for only MD maps, (4) double-input for FA and MD maps, (5) double-input for DTI images and FA maps (6), and double-input for DTI images and MD maps. The network trained by FA and MD maps exhibited poor diagnostic accuracy of 40% and 50%, respectively. However, the best accuracy (67%) and area under the receiver (AUC) (69%) were achieved by the model trained with only DTI as input. Results demonstrated that the models trained with FA and MD maps are not appropriate for early PD detection. Moreover, the results showed that even if these maps are used as an additional image next to the DTI data, they could not improve the results of the model compared to the model trained with only DTI map. The best accuracy and AUC were achieved by the model trained only with DTI map. Though the accuracy obtained from the different models is relatively low, the focus of this study was to determine the best-performing model/scenario for early PD detection.