Deep learning-based analysis of functional MRI and diffusion tensor imaging for Parkinson's disease diagnosis and progression monitoring

Abstract

Parkinson's Disease (PD) refers to the chronic movement disorder caused by the degeneration of the brain's motor functions. Functional Magnetic Resonance Imaging (fMRI) and Diffusion Tensor Imaging (DTI) are evidence-based neuroimaging procedures which evoke relevant anatomical and functional alterations in PD. The purpose of this work is to investigate whether the machine learning approach can benefit from applying deep learning for data interpretation of fMRI and DTI to detect disease at an early stage and study the progression of the disease. The objectives of this research are twofold: first, to predict the biomarkers for attending to the changed brain activity pattern using deep learning model from the fMRI data and the second one is, to find the micro structural changes in the white matter tracts that is specific to the PD using DTI data. The adopted approach is data pre-processing to clean the neuroimaging data and remove different artifacts, then features extraction using deep learning approaches such as CNNs and transformers. Data were collected from the intersection of the PD patients and controls, and similar to the machine learning models, the performance of the segmentation models was assessed using the accuracy, precision, and F1 score based on the databases of PD patients and age-matched healthy controls. Analysis shows that the newly developed deep learning models outperforms previous conventional machine learning techniques with more significant increases in sensitivity for early-stage PD diagnosis. Respective investigation of feature importance provided significant BrainNet features related to PD diagnosis and identified main brain areas and white matter tracts involved in disease, concordant with prior clinical research. To sum up, the findings of the presented work can be useful for developing deep learning algorithms for the analysis of fMRI and DTI data in the context of PD diagnosis and further research. Lastly, the general avenue of future work will cover the combination of multiple modalities and the testing of the models on bigger and more diverse datasets.