ZHANG, HAOZHENG (2024) Human Movement Disorders Analysis with Graph Neural Networks. Doctoral thesis, Durham University.
| PDF 1492Kb |
Abstract
Human movement disorders encompass a group of neurological conditions that cause abnormal movements. These disorders, even when subtle, may be symptomatic of a broad spectrum of medical issues, from neurological to musculoskeletal. Clinicians and researchers still encounter challenges in understanding the underlying pathologies. In light of this, medical professionals and associated researchers are increasingly looking towards the fast-evolving domain of computer vision in pursuit of precise and dependable automated diagnostic tools to support clinical diagnosis. To this end, this thesis explores the feasibility of the interpretable and accurate human movement disorders analysis system using graph neural networks.
Cerebral Palsy (CP) and Parkinson’s Disease (PD) are two common neurological diseases associated with movement disorders that seriously affect patients’ quality of life. Specifically, CP is estimated to affect 2 in 1000 babies born in the UK each year, while PD affects an estimated 10 million people globally. Considering their clinical significance and properties, we develop and examine the state-of-the-art attention-informed Graph Neural Networks (GNN) for robust and interpretable CP prediction and PD diagnosis.
We highlight the significant differences between the human body movement frequency of CP infants and healthy groups, and propose frequency attention-informed convolutional networks (GCNs) and spatial frequency attention based GCNs to predict CP with strong interpretability. To support the early diagnosis of PD, we propose novel video-based deep learning system, SPA-PTA, with a spatial pyramidal attention design based on clinical observations and mathematical theories. Our systems provide undiagnosed PD patients with low-cost, non-intrusive PT classification and tremor severity rating results as a PD warning sign with interpretable attention visualizations.
Item Type: | Thesis (Doctoral) |
---|---|
Award: | Doctor of Philosophy |
Faculty and Department: | Faculty of Science > Computer Science, Department of |
Thesis Date: | 2024 |
Copyright: | Copyright of this thesis is held by the author |
Deposited On: | 16 Apr 2024 12:50 |