Deep Learning-Based Automatic Segmentation of Skeletal Muscles and Generalisation Study

Musculoskeletal disorders affect a significant portion of the population, with one in four people in the UK currently suffering from such conditions, which impact both individuals’ work and social lives. While the mechanisms underpinning skeletal disorders are well-understood, a major challenge in advancing the understanding of muscle-related conditions lies in the difficulty of accurately measuring muscle tissue's physiological status.
Muscle disorders vary significantly in terms of their causes, affected muscles, progression rates, and treatment strategies. Furthermore, individuals with the same muscle disorder often exhibit different responses to the condition, highlighting the need for subject-specific, quantitative characterisation of muscle tissue in vivo. This could significantly enhance current diagnosis and treatment strategies and provide a more informed approach to evaluating the efficacy of new treatments in clinical trials. However, despite its potential, quantitative muscle tissue analysis has not yet been fully integrated into clinical practice. As an indispensable part of this quantitative analysis, manual segmentation of muscle images is labour-intensive, time-consuming, prone to inter- and intra-operator variability, highlighting the need for automatic segmentation methods.
The aim of this thesis was to develop, test, and analyse methods for deep learning based automatic muscle segmentation from medical imaging data. This work presents three distinct methods designed to address the limitations of the current method for muscle automatic segmentation in MR images. The outcome provides a comprehensive overview of both existing and novel methods for muscle segmentation pipeline and analysis from medical images.
The methods discussed in this thesis offer valuable insights for future research, providing a foundation for the quantitative study of muscle segmentation. By adopting the best-suited deep learning models or pre-/post-processing pipeline from this work, future studies can improve the understanding and treatment of muscle conditions. Ultimately, this research aims to promote the clinical adoption of computational tools for muscle disorder characterisation, enhancing diagnosis, treatment planning, and patient monitoring.