Optimizing optical coherence tomography image reconstruction: a comparative study of image generation methods based on diffusion models
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This thesis explores the enhancement of optical coherence tomography (OCT) images using diffusion models, transforming subsampled images into high-resolution versions. the subsampling process is based on three techniques—random, blue noise, and Bayer—at 10% and 25% transmittance to simulate image degradation. These subsampled images are then used as input for the reconstruction models. ☐ The training phase employs Denoising Diffusion Probabilistic Models (DDPM), based on Ho et al., which initially learns from a vast dataset to understand OCT images’ intrinsic properties. The study then applies two advanced methods, RePaint and Diffusion Posterior Sampling (DPS), for reconstructing high-resolution OCT images. RePaint and DPS, developed by Lugmayr et al. and Chung et al. respectively, start with a DDPM’s pre-trained model. ☐ The effectiveness of each mask in retaining critical image details and the performance of RePaint and DPS in reconstructing high-resolution images are evaluated. The study measures how each method impacts image fidelity, providing insights into their practical use in medical imaging. ☐ Results show varying performances across different masks and reconstruction techniques, suggesting tailored strategies for specific OCT scenarios. This research enhances medical imaging by offering comparisons and insights that could improve diagnostic accuracy and patient outcomes.
Description
Keywords
Deep learning, Diffusion models, Optical coherence tomography, Diffusion Posterior Sampling, High-resolution images