Self-supervised speckle noise reduction of optical coherence tomography without clean data

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computer Vision

Background:

• Optical Coherence Tomography (OCT) is vital for non-invasive, high-resolution imaging in clinical diagnosis.
• Speckle noise in OCT images degrades quality and can obscure fine details, impacting diagnostic accuracy.
• Supervised deep learning methods for denoising require extensive paired noisy-clean image datasets, which are challenging to obtain in clinical settings.

Purpose of the Study:

• To develop a self-supervised deep learning strategy for OCT image despeckling.
• To overcome the limitation of requiring paired noisy-clean images for training denoising models.
• To improve OCT image quality for enhanced clinical diagnosis.

Main Methods:

• A self-supervised learning framework was proposed, training a deep neural network using a single noisy OCT image.
• Adjacent pixel patches were randomly sampled to create paired undersampled input and target images for network training.
• A multi-scale pixel patch sampler and specialized loss functions were employed to ensure effective despeckling and structure preservation.

Main Results:

• The proposed self-supervised method demonstrated superior performance in both despeckling and structure preservation compared to state-of-the-art techniques.
• Quantitative evaluations and qualitative visual comparisons validated the method's effectiveness.
• The approach significantly simplifies the training and deployment process by eliminating the need for clean OCT images.

Conclusions:

• The self-supervised strategy offers a practical and efficient solution for OCT image despeckling in clinical practice.
• This method enhances image quality without compromising important structural information, aiding in more reliable diagnoses.
• The ease of training and deployment makes this technique highly valuable for real-world clinical applications.