Automated quality control of temporalis muscle segmentations for sarcopenia screening using magnetic resonance imaging in the UK biobank
File(s)
Author(s)
Wang, James
Type
Thesis
Abstract
Skeletal muscle is a dynamic tissue that responds to ageing, exercise, and disease through
hypertrophy or atrophy. Age-related sarcopenia, defined as low muscle mass and function,
is typically assessed on appendicular muscle. In contrast, cancer-associated muscle loss is
often evaluated on truncal and abdominal regions using opportunistic assessment from routine
imaging. A scoping review revealed significant heterogeneity in methodology, terminology,
and anatomical focus, highlighting inconsistencies in biomarker application. Moreover, the
relationship between age-related changes in truncal versus appendicular muscle remains poorly
defined.
Using data from the UK Biobank (UKBB), this thesis derives sex-specific T-score thresh-
olds for skeletal muscle cross-sectional area at the third lumbar vertebra (L3), a commonly
used radiological biomarker. These are correlated with standard sarcopenia metrics—DEXA,
bioimpedance analysis, and hand grip strength—bridging imaging-based truncal biomarkers
with traditional appendicular assessments and addressing a key gap in sarcopenia research.
To enable scalable clinical deployment of automated muscle quantification, the thesis evaluates
Reverse Classifier Accuracy (RCA), an unsupervised segmentation quality assurance method.
Implemented within U-Net and DeepMedic architectures, and optimised using transfer learning,
RCA reduces overfitting and improves segmentation robustness. When applied to temporalis
muscle segmentation in UKBB head MRI, RCA outperformed heuristic thresholds in identi-
fying poor-quality segmentations across both sexes. Although temporalis muscle area did not
significantly correlate with frailty metrics in this cohort, RCA emerged as a reliable, scalable
solution for quality control in automated workflows.
The integration of transfer learning into RCA constitutes a novel advance in the evaluation
of radiological segmentations. This thesis lays the groundwork for autonomous, opportunistic
sarcopenia screening in routine imaging. Future validation studies—particularly in glioma
populations—are warranted to assess the prognostic potential of temporalis muscle biomarkers
and ultimately support earlier clinical interventions.
hypertrophy or atrophy. Age-related sarcopenia, defined as low muscle mass and function,
is typically assessed on appendicular muscle. In contrast, cancer-associated muscle loss is
often evaluated on truncal and abdominal regions using opportunistic assessment from routine
imaging. A scoping review revealed significant heterogeneity in methodology, terminology,
and anatomical focus, highlighting inconsistencies in biomarker application. Moreover, the
relationship between age-related changes in truncal versus appendicular muscle remains poorly
defined.
Using data from the UK Biobank (UKBB), this thesis derives sex-specific T-score thresh-
olds for skeletal muscle cross-sectional area at the third lumbar vertebra (L3), a commonly
used radiological biomarker. These are correlated with standard sarcopenia metrics—DEXA,
bioimpedance analysis, and hand grip strength—bridging imaging-based truncal biomarkers
with traditional appendicular assessments and addressing a key gap in sarcopenia research.
To enable scalable clinical deployment of automated muscle quantification, the thesis evaluates
Reverse Classifier Accuracy (RCA), an unsupervised segmentation quality assurance method.
Implemented within U-Net and DeepMedic architectures, and optimised using transfer learning,
RCA reduces overfitting and improves segmentation robustness. When applied to temporalis
muscle segmentation in UKBB head MRI, RCA outperformed heuristic thresholds in identi-
fying poor-quality segmentations across both sexes. Although temporalis muscle area did not
significantly correlate with frailty metrics in this cohort, RCA emerged as a reliable, scalable
solution for quality control in automated workflows.
The integration of transfer learning into RCA constitutes a novel advance in the evaluation
of radiological segmentations. This thesis lays the groundwork for autonomous, opportunistic
sarcopenia screening in routine imaging. Future validation studies—particularly in glioma
populations—are warranted to assess the prognostic potential of temporalis muscle biomarkers
and ultimately support earlier clinical interventions.
Version
Open Access
Date Issued
2025-07-30
Date Awarded
2026-05-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Williams, Matthew
McGregor, Alison
Publisher Department
Department of Surgery & Cancer
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)