Deep self-learning enables fast, high-fidelity isotropic resolution restoration for volumetric fluorescence...

Area of Science:

• Microscopy
• Image Processing
• Computational Biology

Background:

• 3D fluorescence microscopy suffers from resolution anisotropy, limiting 3D image quality and analysis.
• Lateral and axial resolution differences hinder accurate reconstruction and interpretation of biological structures.

Purpose of the Study:

• To introduce Self-Net, a deep self-learning method to overcome resolution anisotropy in 3D fluorescence microscopy.
• To significantly improve axial image resolution using lateral images from the same dataset.

Main Methods:

• Leveraging natural anisotropy for deep self-learning.
• Incorporating unsupervised learning for anisotropic degradation simulation.
• Utilizing supervised learning for high-fidelity isotropic image recovery.

Main Results:

• Self-Net effectively suppresses hallucination and enhances image quality.
• Reconstructed high-fidelity isotropic 3D images across organelle to tissue scales.
• Enabled first-time isotropic whole-brain imaging at 0.2μm voxel resolution.

Conclusions:

• Self-Net addresses the critical issue of resolution anisotropy in 3D fluorescence microscopy.
• The method offers a cost-effective solution for high-quality 3D biological imaging.
• Self-Net is a versatile approach applicable to diverse microscopy platforms and biological samples.