Attention-based Deep Neural Network for Early Detection of Parkinson's disease using Diffusion Tensor Imaging

Accurate and early diagnosis of Parkinson's (PD) disease is a compelling need due to the lack of motor symptoms. In previous studies, statistical analysis and support vector machine-based (SVM) classification performed on fractional anisotropy (FA) and mean diffusivity (MD) maps extracted from diffusion tensor imaging (DTI) images showed significantly reduced fractional anisotropy (FA) and increased mean diffusivity (MD) in many brain regions of PD group compared to healthy control. Though the diagnostic accuracy of PD based on statistical analysis and SVM of DTI images are relatively good, convo-lutional neural networks (CNNs) are the leading techniques for classifying patients into PD and healthy. However, conventional CNNs generally use the whole image as input for network learning, while in some cases, the focus would be only on specific regions within the image. In this work, we propose an attention-based neural network using a 3D ResNet-18 model with an additional compartment, which converts the region of interest (ROI) masks into an attention map to diagnose PD from diffusion tensor imaging (DTI). In this regard, the focus would be only on the specific brain regions affected by PD. The accuracy and area under the receiver operating characteristic (ROC) with and without attention module were 67%, 83%, and 69%, 84%, respectively. The results demonstrated that the attention-based model exhibited noticeably superior performance owing to the concentration on the key regions of the brain. Therefore, the attention-based 3D ResNet-18 can reduce the number of misdiagnoses as a decision support tool for neurologists.