Mfmdenoiser: Self-Supervised Denoising for Structured Noise in Fluorescence Microscopy Image Stacks Via Orthogonal Masks

Fluorescence microscopy is essential for detecting and tracking certain items inside cells in biological research. However, complex noise is inevitable during image capture due to optical characteristics and hardware limitations. Current self-supervised methods often model the noise as Poisson-Gaussian noise, overlooking potential structured noise that may violate the pixel-wise independence assumption. To address this, we propose a Multi-slice Fusion Module (MFM) and combine it with the Blind-spot Denoising Network (BSD) to provide a self-supervised denoising algorithm. MFM leverages the inter-slice continuity of fluorescence microscopy stacks and uses orthogonal masks to fuse multiple slices of stacks, breaking the spatial correlation of noise while maintaining image content. We have validated the effectiveness of our method on datasets containing structured noise and discussed its variant and implementation details. The code is accessible at https://github.com/chill-luo/MFMdenoiser.