Kok, Yong En
(2025)
Deep tissue analysis: advancing optical techniques with interpretable deep learning and aberration correction.
PhD thesis, University of Nottingham.
Abstract
The ability to visualise and analyse biological tissues at the molecular and cellular levels plays a critical role in early disease diagnosis and improved patient outcomes. Optical technologies offer rapid, cost-effective diagnostic capabilities with minimal invasiveness, yet their practical application is hindered by noise, data dependencies, and day-to-day setup variations. In response, this thesis explores how deep learning can elevate optical technologies along two complementary frontiers: (i) classification of Musculoskeletal (MSK) diseases with Raman spectra, and (ii) high-resolution tissue imaging through Adaptive Optics (AO) for correcting severe optical aberrations. Both tasks emphasise robust, interpretable frameworks, acknowledging that model performance must be supported by domain knowledge and transparency if it is to be trusted in clinical workflows.
Firstly, the thesis identifies spectral noise and heavy reliance on manual pre-processing as major barriers to deploying Raman spectroscopy for MSK disease diagnosis. To address this, a deep learning pipeline is introduced to automatically learn the optimal spectral pre-processing strategies for accurate classification of Osteoarthritis and healthy cartilage, thereby reducing human subjectivity and processing time. Using explainable Artificial Intelligence methods, the thesis validates that the network's decisions align with known biomarkers of disease. The work further demonstrates that classification with comparable accuracy can be achieved using only a minimal subset of these key spectral features, potentially accelerating data acquisition without compromising diagnostic accuracy.
Secondly, this thesis investigates deep learning solutions for optical aberration correction in high-resolution tissue imaging. The work systematically examines optimal input requirements for aberration correction, particularly for large aberrations encountered in deep tissue imaging. To address day-to-day experimental variations, the thesis explores transfer learning strategies that reduce the need for extensive training data while maintaining correction accuracy. Building upon these findings, the thesis further develops a novel physics-informed graph neural network that explicitly models the underlying relationships among Zernike polynomials while enforcing consistency between predicted aberrations and image content through a frequency-aware alignment loss. This state-of-the-art approach not only yields accurate and physically valid image restorations across diverse samples and imaging conditions but also provides model-specific explanations through Zernike coefficient predictions, offering interpretable insights into optical aberrations.
In summary, this thesis advances optical diagnostic technologies from theoretical demonstrations toward clinical feasibility. By addressing data dependencies, noise sensitivity, and interpretability, this research preserves the inherent advantages of optical systems while harnessing deep learning to enhance reliability, accessibility, and trustworthiness for biomedical applications in response to growing healthcare demands.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[thumbnail of Correction]](https://eprints.nottingham.ac.uk/style/images/fileicons/application_pdf.png "Correction")