Volumetric fluorescence microscopy often suffers from anisotropic spatial resolution, where axial resolution is inferior to lateral resolution. This paper presents a deep-learning-based unsupervised super-resolution technique to enhance anisotropic images in volumetric fluorescence microscopy. Unlike existing methods requiring matched high-resolution images for training, this method uses only a single 3D image stack. It employs an optimal transport-driven cycle-consistent generative adversarial network (OT-cycleGAN) to learn from unpaired high-resolution 2D lateral images and low-resolution 2D images from other planes. Experiments using confocal and light-sheet microscopy demonstrate enhanced axial resolution, restoration of visual details, and removal of imaging artifacts.

Hyoungjun Park, Myeongsu Na, Bumju Kim, Soohyun Park, Ki Hean Kim, Sunghoe Chang, Jong Chul Ye