Fig. 1: Principle of LF-denoising.

a Comparisons between strategies of self-supervised denoising by relying on temporal, spatial, and spatial-angular redundancy. b Schematic of the LF-denoising processing pipeline. Noisy spatial-angular images are first converted into two EPIs via bidirectional angular traversing. These noisy EPIs are divided into four sub-stacks along the x and y dimensions, serving as inputs and targets for training transformer-based networks in a self-supervised manner. An attention-based module then fuses the two channels to produce clean spatial-angular images. Multiple orthogonal masks and downsampling operators are employed to establish a global loss function, enhancing fidelity. Loss 1 and Loss 2 are sub-network loss functions defined on the EPI representation, whereas Loss 3 is a global loss function defined on fused spatial-angular representation. c Spatial-angular images of simulated neuron data before and after LF-denoising, with clean images provided as ground truth. SNR values are indicated above each image. LF-denoising performance indicated by the SSIM across varying photon numbers (d) and Gaussian standard deviations (e), with distinct improvement over state-of-the-art methods. Image bit depth is set to 16. Scale bars: 20 μm (b, c).