End-to-end design for visible light OCT denoising by speckle reduction scanning and deep learning

Abstract

Purpose : Visible light optical coherence tomography (VIS-OCT) is an emerging imaging modality that provides one-micron level axial resolution. Improving VIS-OCT image quality by denoising is an essential step in the overall workflow in VIS-OCT clinical applications. In this study, we provide the first deep learning-based, end-to-end design for VIS-OCT denoising using our speckle reduction scanning dataset.

Methods : The HD dataset includes 105 retinal B-scans obtained on our 2nd Gen dual-channel VIS-OCT system for 12 subjects. With our speckle reduction scanning protocol, 16 or 32 B-scans at the same location were obtained and averaged to get the noisy-clean pairs. We provided both supervised and self-supervised (based on N2V) strategies to fulfill practical scenarios where clean images are not available. We splited the HD dataset into training, validation, and test sets as (51:17:37). The dataset was augmented 8-fold by rotations and horizontal flip. The images were resized to 512x512 as the input. We used the U-Net architecture and trained the models for 200 epochs by Adam optimizer. We also tested the models on our Raster scan dataset with a 3D volume of the human retina.

Results : Fig1. A-D visualize the denoising performance on HD dataset. We repeated the experiment 8 times and the average PSNR/SSIM for self-supervised and supervised strategies are 27.26(±0.2470)/0.5296(±0.0083) and 30.96((±0.1556)/0.7636((±0.0024), respectively, both of which significantly improved the PSNR/SSIM of noisy images from 22.90/0.2639.

Fig. 2 A-C qualitatively show the models trained on HD dataset generalize well to Raster scan dataset and denoise the whole 3D volumes.

Conclusions : The proposed end-to-end framework that includes both self-supervised and supervised strategies successfully denoise the HD VIS-OCT images with significantly improved PSNR and SSIM. The framework is also robust to denoise the never-seen Raster scan VIS-OCT volume. This study reveals the strength, robustness and flexibility of deep learning-based denoising for VIS-OCT images.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Fig .1. A) A noisy example from the HD testing dataset. B) The denoised result of the self-supervised strategy. C) The denoised result of the supervised strategy. D) The multi B-scans averaged clean image.

 

Fig. 2. Top: B-scan, bottom: en face image. Fig. 2A, B and C are the noisy volume, volumes denoised by the self-supervised method and the supervised method.