Self-supervised Self2Self denoising strategy for OCT speckle reduction with a single noisy image.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is susceptible to speckle noise due to its low-coherence interferometry.
• Existing deep learning methods for OCT despeckling often require ground-truth images, which are challenging to obtain in clinical settings.
• Speckle noise degrades the quality and diagnostic accuracy of OCT images.

Purpose of the Study:

• To develop a self-supervised deep learning scheme for OCT despeckling that does not require ground-truth images.
• To reduce the impact of speckle noise in OCT images using a single noisy input.
• To enhance the performance of OCT image analysis through effective noise reduction.

Main Methods:

• A self-supervised deep learning strategy named Self2Self strategy (S2Snet) is proposed.
• The S2Snet architecture utilizes a gated convolution layer to update its partial convolution.
• The network processes both the input image and its Bernoulli sampling instances, incorporating a custom loss function to eliminate background noise and averaging multiple predictions for the final denoised output.

Main Results:

• The proposed S2Snet scheme demonstrates superior performance in OCT despeckling compared to existing methods.
• S2Snet outperforms other self-supervised deep learning techniques and traditional non-deep learning methods.
• Quantitative improvements include a 3.41% increase in Peak Signal-to-Noise Ratio (PSNR) and a 2.37% increase in Structural Similarity Index Measure (SSIM) compared to the original Self2Self network, and significant gains over the NWSR method (19.9% PSNR, 22.7% SSIM).

Conclusions:

• The S2Snet method offers an effective solution for OCT despeckling without relying on ground-truth data.
• This self-supervised approach significantly improves image quality in OCT, making it valuable for clinical applications.
• S2Snet represents a promising advancement in noise reduction techniques for medical imaging.