Motion-corrected Noise2Noise denoising for structural analysis of cells and tissues with (cryo-)FIB/SEM

Abstract

MOTIVATION: (Cryo-)FIB/SEM enables three-dimensional (3D) ultrastructural analysis of cells and tissues at nanoscale resolution. However, its characteristically low signal-to-noise ratio severely hinders interpretation of the volume data. Noise2Noise has emerged as a state-of-the-art denoising paradigm with outstanding performance in 3D electron microscopy -particularly cryo-ET- but its direct application to (cryo-)FIB/SEM remains unfeasible due to the inherent acquisition workflows. RESULTS: A method to enable Noise2Noise denoising in (cryo-)FIB/SEM is introduced. The input volume is split into even and odd slices followed by the application of motion correction methods to estimate and compensate for the structural differences between the two subvolumes. The resulting motion-corrected subvolumes are then used to train the Noise2Noise model, which is ultimately applied to denoise the original volume. Representative (cryo-)FIB/SEM volumes have been used to validate the method and to demonstrate the critical importance of the motion correction stage for volumes with a slice thickness greater than 10 nm. AVAILABILITY AND IMPLEMENTATION: Code is available at https://github.com/microscopy-processing/N2N-REO and https://doi.org/10.5281/zenodo.22303478. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

The authors' abstract, as published at the source. Bioinformatics, 2026 · DOI ↗