Domain-Aware Few-Shot Learning for Optical Coherence Tomography Noise Reduction.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Imaging

Background:

• Speckle noise is a persistent challenge in medical imaging, particularly in Optical Coherence Tomography (OCT).
• Deep learning methods have advanced noise reduction, but supervised models struggle with domain shifts caused by varying acquisition parameters and biological tissues.
• Existing deep neural networks (DNNs) are vulnerable to changes in sampling space, resolution, and contrast, degrading performance on unseen data.

Purpose of the Study:

• To develop a few-shot supervised learning framework for efficient OCT noise reduction.
• To address the domain shift problem in OCT imaging systems.
• To enable high-speed training and adaptation of despeckling models with minimal data.

Main Methods:

• Proposed a few-shot supervised learning framework for OCT noise reduction.
• Developed a method for formulating and addressing domain shift issues in OCT.
• Implemented and compared practical variations of the proposed framework for robustness and efficiency.

Main Results:

• Achieved high-speed training (seconds) using only a single image and its ground truth.
• Demonstrated that output resolution of despeckling models is linked to source domain resolution.
• Showcased improved sample complexity, generalization, and time efficiency for noise reduction models.

Conclusions:

• The proposed few-shot learning framework significantly enhances adaptability and efficiency for OCT noise reduction.
• The study provides insights into domain shift challenges and potential solutions for OCT imaging.
• The framework shows potential for real-time computer vision applications beyond medical imaging.