Self-supervised denoising of light field microscopy for high-dimensional neural imaging

Abstract

Photon noise arising during image acquisition remains a major obstacle to resolving fine structures in optical microscopy, particularly under low-light conditions. Due to the inherent difficulty to obtain noise-free ground truth data, selfsupervised denoising approaches have been widely adopted for microscopy imaging. However, previous self-supervised methods rely on either temporal redundancy across frames or spatial redundancy within single images, which leads to the degradation in temporal or spatial resolution on high dynamic neuron imaging. Here, we proposed a self-supervised denoising network specifically tailored for light field neural imaging data. The proposed network incorporates a tworoute structure with different light field epipolar plane realignment. We further design a spatial angular attention model to extract high-dimensional contexts in spatial-angular light field measurements without relying solely on temporal or spatial information. We demonstrated that our method preserves high fidelity and maintains high resolution in extreme low light dose condition, highlighting its potential for robust, low-phototoxicity neuroscience applications.