Deep Learning Methods for Muscle Analysis From Magnetic Resonance Images

Quantitative analyses of skeletal muscles inform diverse fields, from assessing training impacts in sports science to muscular diseases in medicine, to name a few. Medical imaging offers noninvasive tools for muscle size and structure analysis, typically necessitating pixel/voxel-wise annotations. Segmentation frameworks utilising deep learning offer a data-driven approach to automate the annotation process efficiently and accurately. However, their performance is typically dependent on the nature, distribution and size of the training data. Four studies are presented in this thesis to enable automatic segmentation of skeletal muscles in various scenarios. The first study systematically evaluated the efficacy of 2D, 3D, and hybrid models in segmenting skeletal muscles from anisotropic magnetic resonance imaging (MRI) scans and concluded that the hybrid model demonstrated the best performance. This finding is used in the second study, where a novel hybrid model is proposed. It also demonstrated human-level performance on isotropic MRI scans, showing the potential to automate the annotation process for a wide range of data. The third study focuses on medical image segmentation in the semi-supervised setting and proposes a hybrid model based on the mean-teacher network. The proposed hybrid model, which we named HD-Teacher, conducts segmentation with hybrid features extracted from 2D and 3D dimensions. In addition, it requires minimal labelled training data and can utilise unlabelled data to improve its performance through a novel hybrid regularisation mechanism. Its practicality is further demonstrated through its state-of-the-art performance on musculoskeletal data. We also further evaluate its performance on two other datasets with varying anisotropy to showcase its generalisability. The fourth study focuses on domain adaptation to overcome the challenge that most deep learning models trained on one specific type of image (source domain) typically do not perform well on other types of images (target domain). To overcome this challenge, a continual test-time domain adaptation method is proposed to first train the model in the source domain with domain-generalised data augmentation methods to ensure an adequate initial performance on the target domain. Then, a mean-teacher network is used to improve the model’s performance in the target domain through consistency regularisation and a novel uncertainty-ranked cross-task regularisation. The proposed framework achieves state-of-the-art performance on cross-domain muscle and four other similar datasets, showing its generalisability and the potential to enable large-scale quantitative analysis of medical data for a wide range of anatomical structures and in a clinical setting. In summary, this thesis presents four studies that introduce deep learning methods tailored for medical image segmentation across different scenarios. The proposed methods demonstrate significant technical novelty. More importantly, they enable large-scale quantitative medical research in various conditions by reducing the cost and subjectivity associated with manual labelling. Finally, this thesis provides pointers for potential directions for future development of the presented methods to improve their performance and practicality further.

Grantor dc:publisher

UNSW, Sydney

Year dc:date

2024

Author dc:creator

• Zhu, Jiayi ; https://orcid.org/0000-0003-3837-7631

• Computer Vision
• Deep Learning
• Medical Image Analysis
• MRI
• CT
• anzsrc-for: 461103 Deep learning
• anzsrc-for: 4603 Computer vision and multimedia computation
• anzsrc-for: 460306 Image processing
• anzsrc-for: 4611 Machine learning

Statement dc:rights

• open access
• CC BY 4.0
• free_to_read

Licence

• https://purl.org/coar/access_right/c_abf2
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

en

Identifier

https://doi.org/10.26190/unsworks/30447

OAI identifier oai:identifier

oai:unsworks.library.unsw.edu.au:1959.4/102882