A General Computational Framework for Restoring Cellular-Resolution in Optical Coherence Tomography Images of the Eye

Abstract

Achieving full-volume cellular-resolution imaging of the eye with Optical Coherence Tomography (OCT) is constrained by aberrations that limit the depth range over which sufficient lateral resolution and Signal-to-Noise Ratio (SNR) can be maintained. Among many solutions proposed for this problem, computational methods stand out as they offer the correction of these aberrations without compromising imaging speed or quality, and without introducing additional costs to the OCT setups. However, these solutions are only applicable to the data acquired with complex OCT systems featuring high phase stability. In addition, the in vivo application of these solutions in the human eye has been limited to the posterior layers, i.e., retinal tissues. The proposed research aims to address these challenges through innovative computational methods that do not rely on phase-stable OCT setups. Such development holds the potential to offer a system-agnostic approach for cellular resolution imaging of the human eye in full volumes, enabling seamless integration into any clinical OCT system for diagnostic purposes.

This thesis proposes two computational methods for enhancing image reconstruction in cellular-resolution OCT. First, aberration- and noise-compensated reconstruction is addressed using non-local image priors within a Maximum a Posteriori (MAP) framework, yielding improved cellular-level detail across various tissue types, including the healthy human cornea. Building on this foundation, a Physics-Informed Diffusion Model (PIDM) is developed for super-resolved and defocus-corrected OCT image reconstruction. The model integrates non-local priors into the diffusion process, enabling conditioning of inputs and latent variables that guide reverse diffusion. By leveraging both the flexibility of diffusion models and the physical principles of OCT imaging, PIDM achieves high-fidelity correction of image degradations. Experimental validation against well-established ground truths demonstrates that PIDM outperforms baseline methods, including the MAP framework in the first project and a standalone super-resolution diffusion model. Quantitative evaluations highlight improvements in SNR, sharpness, and contrast, while qualitative assessments demonstrate its potential to facilitate key applications such as cell counting, segmentation, and dynamic cell studies in the human cornea.

By addressing key limitations in cellular-resolution OCT imaging of the eye, this research has the potential to enable non-invasive, high-fidelity cellular imaging of the eye in full volumes. The proposed methods could facilitate widespread clinical adoption of cellular-resolution OCT, extending its diagnostic capabilities for early detection of eye diseases.