Deep learning enables reference-free isotropic super-resolution for volumetric fluorescence microscopy.

Area of Science:

• Biomedical Imaging
• Microscopy
• Artificial Intelligence

Background:

• Volumetric fluorescence microscopy often suffers from anisotropic spatial resolution, limiting axial detail.
• Existing super-resolution methods typically require matched high-resolution target images for training.

Purpose of the Study:

• To present a deep-learning-enabled unsupervised super-resolution technique for enhancing anisotropic volumetric fluorescence microscopy images.
• To reduce the practical effort required for super-resolution imaging by eliminating the need for paired training data.

Main Methods:

• Developed an optimal transport-driven cycle-consistent generative adversarial network (GAN).
• The network trains on unpaired 2D image sets from different resolution planes within a single 3D image stack.
• No prior knowledge of image formation, data registration, or separate target data acquisition is needed.

Main Results:

• Successfully enhanced axial resolution in volumetric fluorescence microscopy.
• Restored suppressed visual details between imaging planes.
• Effectively removed common imaging artifacts.
• Demonstrated efficacy using fluorescence confocal and light-sheet microscopy.

Conclusions:

• The unsupervised deep learning approach significantly improves volumetric fluorescence microscopy resolution.
• This method simplifies super-resolution implementation by removing data acquisition and registration burdens.
• The technique offers a practical solution for enhancing anisotropic image quality in 3D microscopy.