Abstract
Purpose :
Optical Coherence Tomography (OCT) is essential in eye care but is limited by speckle noise, which obscures important details. Our study developed a self-supervised framework for OCT image denoising, improving clarity without needing repeated scans, clean references, or extensive preprocessing.
Methods :
This research refines the self-supervised noise2noise framework (Lehtinen, ICML, 2018), traditionally reliant on multiple noisy instances of identical target images for denoising. Inspired by Gisbert's ARVO 2020 research, which employed template-based spatial alignment of repeated OCT scans, our study acknowledges the practical difficulty in acquiring such scans. Consequently, we exploit the structural similarity between adjacent B-scans in an OCT volume, considering them as separate noisy representations of the same underlying tissue structure.
A principal challenge in implementing noise2noise is image alignment, conventionally a time-intensive pre-processing stage. Our framework incorporates an image registration module (Balakrishnan, CVPR, 2018), enabling seamless end-to-end training. This integration not only streamlines the data preparation but also markedly reduces the time required for aligning B-scan slices.
For training and testing, our dataset comprises Cirrus HD-OCT ONH scans, with a voxel resolution of 200x1024x200 over dimensions of 6x2x6 mm3. It includes 20 scans for training and 9 for testing, establishing a solid base for evaluating our denoising approach.
Results :
We used synthetic data to mimic speckle noise, lacking clean reference images. Image quality was evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). Higher PSNR reflects better image quality, while SSIM assesses visual aspects like luminance and contrast. Our approach achieved a PSNR of 25.0 and an SSIM of 0.390, exceeding BM3D (PSNR: 23.3, SSIM: 0.272) and Self Fuse (PSNR: 21.4, SSIM: 0.231). This indicates superior denoising, as also evident in Figure 1.
Conclusions :
Our self-supervised denoising framework enables the training of a denoising network without the need for repeated scans, clean targets, or extensive preprocessing. It exhibits significant qualitative and quantitative improvements in OCT scan quality, underscoring its potential in enhancing ophthalmic diagnostics.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.