Zero-Shot Enhancement With Cross-Modal Applicability for Low-Light Vis-$\mu$OCT Images.

Area of Science:

• Biomedical Imaging
• Optical Engineering
• Computer Vision

Background:

• Optical coherence tomography (OCT) provides rapid, non-destructive imaging but suffers from decreased brightness in deep tissues or under low power.
• Visible-light micro-OCT (vis-μOCT) is particularly limited by scattering and penetration depth due to shorter wavelengths.
• Insufficient backscattered light significantly restricts the application of OCT in challenging imaging scenarios.

Purpose of the Study:

• To introduce DifNIR, a novel framework for enhancing low-light OCT images.
• To address the issue of image brightness degradation in OCT, especially in vis-μOCT.
• To improve the penetration depth and clarity of OCT imaging for deeper tissues.

Main Methods:

• The DifNIR framework incorporates a preliminary denoising stage.
• Image enhancement is achieved using a neural implicit representation (NIR) network with pixel values as auxiliary input.
• Unsupervised learning is enabled through custom-designed loss functions.

Main Results:

• DifNIR demonstrated superior performance on an en face image dataset, achieving high SNR (58.99 dB) and CNR (49.56 dB).
• The method significantly improved visual quality and image metrics, with a NIQE score of 9.0553.
• DifNIR showed strong generalizability, effectively enhancing OCT B-scan and retinal images from different devices.

Conclusions:

• The proposed DifNIR network effectively mitigates brightness degradation, producing clearer and better-illuminated OCT images.
• The framework exhibits strong generalization capabilities across various OCT imaging modalities and datasets.
• DifNIR enables the revelation of deep-layer information, expanding OCT applications into cost-effective, high-speed imaging settings.