Self-supervised Blind2Unblind deep learning scheme for OCT speckle reductions.

Area of Science:

• Biomedical Imaging
• Medical Image Analysis
• Deep Learning

Background:

• Optical coherence tomography (OCT) is crucial for non-invasive imaging.
• Speckle noise in OCT degrades image quality and diagnostic accuracy.
• Existing speckle reduction methods have limitations like high computational cost or reliance on clean image priors.

Purpose of the Study:

• To propose a novel self-supervised deep learning scheme for OCT speckle reduction.
• To develop a method that effectively reduces speckles using only a single noisy OCT image.
• To improve the clinical applicability of OCT by enhancing image quality.

Main Methods:

• Introduced the Blind2Unblind network with refinement strategy (B2Unet).
• Designed a global-aware mask mapper and a re-visible loss function to address network blind spots.
• Utilized a self-supervised learning approach, requiring only noisy OCT images.

Main Results:

• B2Unet effectively suppresses speckles while preserving crucial tissue microstructures.
• Achieved superior performance compared to state-of-the-art model-based and fully supervised deep learning methods.
• Demonstrated robustness across different OCT image datasets and clinical scenarios.

Conclusions:

• B2Unet offers an effective and robust solution for OCT speckle reduction.
• The proposed self-supervised deep learning scheme enhances OCT image quality without prior clean images.
• This method has the potential to significantly improve OCT-based disease diagnosis and clinical applications.