Deep Self-learning Enables Fast, High-fidelity Isotropic Resolution Restoration For Volumetric Fluorescence Microscopy

This article presents a two-stage self-learning strategy for Self-Net, a network that enhances the axial resolution of 3D data. The strategy involves slicing 3D data into 2D image sets containing multiple lateral and axial planes, and then using unpaired lateral and axial images retrieved from the same anisotropic volume to learn the axial-to-lateral mappings through unsupervised training. The second stage of the strategy involves using blurred axial images generated by the first-stage network combined with pixel-aligned HR lateral images to train a supervised network to impose a strong constraint on the isotropic restoration results.