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

Abstract

Optical coherence tomography (OCT) is a rapid and non-destructive imaging technique, but image brightness decreases when imaging deep tissues or under low power and short exposure due to insufficient backscattered light. This issue is more pronounced in visible-light micro-OCT (vis-$\mu$OCT), where shorter wavelengths increase scattering and limit penetration, restricting its application. In this paper, we propose Dif-NIR, a novel framework for enhancing low-light OCT images. The framework begins with a preliminary denoising stage. Image enhancement is then performed using a neural implicit representation (NIR) network, in which pixel values are incorporated as auxiliary input to mitigate the oversmoothing effect of fully connected layers. To enable unsupervised learning, custom-designed loss functions is employed. The proposed method is validated through qualitative and quantitative comparisons on a self-collected en face image dataset. To further assess its generalizability, we also performed experiments on B-scan images and retinal images acquired from other OCT devices. On the en face image dataset, Dif-NIR outperforms existing methods in terms of visual quality, SNR (58.99 dB), CNR (49.56 dB), and NIQE (9.0553). It also effectively generalizes to OCT B-scan images and retinal images acquired by other devices. The proposed network effectively mitigates unpredictable brightness degradation, producing clearer and better-illuminated images while exhibiting strong generalization capability. The network effectively reveals deep-layer information in OCT images and can be applied to expand its usage scenarios to cost-effective and high-speed imaging settings.