Differential Diagnosis and Progression Prediction of Cognition Impairment in Patients with Parkinson Disease by Using Diffusion Magnetic Resonance Imaging in Deep Learning

Cognitive impairment can be observed in patients with Parkinson Disease, which can continue to decline over the disease course and is often associated with increased burden on the patient care and inferior clinical outcome. Magnetic Resonance Image was often prescribed to rule out concomitant neurological order, with the advantage that it can be objective without the requirement of patient collaboration. Therefore, to increase the diagnostic confidence of the clinician, we proposed to investigate its performance as an image based biomarker.The hypothesis is that the cognitive decline might be accompanied with white matter changes in the brain in patients with Parkinson Disease, which can be detected by using diffusion MRI. Furthermore, patients with different extents of cognitive impairment can involve different parts of the brain. Because of different topological features, it is possible to design a machine learning/deep learning based algorithm for the diagnosis of cognitive decline in patients with Parkinson Disease. The aim in this 3 year project is to develop an objective image-based biomarker for the differential diagnosis of cognitive decline in patients with Parkinson Disease. Specifically, we aimed to assess the diagnostic performance of diffusion MRI on patients of Parkinson Disease with normal cognition, mild cognitive impairment and dementia. Secondly we will compare the difference in diffusion MRI between those who are stable from those who progress over a follow-up period of 18 months. Finally we will explore the differential white matter involvement in disease subtypes by using a novel technique, fixel based analysis. We will use convolutional neural network to speed up the image acquisition by reconstructing the diffusion tensor using fewer number of diffusion weighted images. Features will be extracted from the brain parcellated according to human brainnetome, the number of which will be reduced to within a reasonable range by using least absolute shrinkage and selection operator. Models for differential diagnosis and progression prediction will be further evaluated by using convolutional neural network and greedy algorithm and validated by using additional blind data collected from Parkinson's Progression Markers Initiative.We expected that the result can significantly increase the value of a standard neuro-imaging examination, enhance the diagnostic confidence of clinicians, reduce the cost due to fluctuation in patient condition and avoid repeated examinations. The proposal is unique in a sense by combining two novel techniques: fixel based analysis and artificial intelligence.

Project ID:PC10907-0995
External Project ID:MOST109-2314-B182-021-MY3