Diffusion-based deep learning method for augmenting ultrastructural imaging and volume electron microscopy.

Area of Science:

• Cellular and Molecular Biology
• Microscopy and Imaging Technologies
• Computational Biology and Bioinformatics

Background:

• Electron microscopy (EM) and volume EM (vEM) are crucial for visualizing cellular ultrastructure in 3D.
• Current vEM techniques face limitations in imaging speed and quality, restricting the analysis of large biological volumes.
• Achieving isotropic imaging for large-scale vEM remains a significant challenge.

Purpose of the Study:

• To develop a novel algorithmic suite, EMDiffuse, to enhance EM and vEM capabilities.
• To improve the resolution, speed, and isotropic reconstruction of vEM data.
• To enable accurate 3D nanoscale ultrastructure analysis within large biological volumes.

Main Methods:

• Leveraged cutting-edge image generation diffusion models for EM and vEM data processing.
• Developed EMDiffuse algorithms for denoising, super-resolution, and isotropic vEM reconstruction.
• Fine-tuned algorithms using minimal data (one pair of 3-megapixel images) for robust transferability.

Main Results:

• EMDiffuse generates realistic, high-resolution ultrastructural details.
• The algorithms demonstrated robust transferability across different vEM techniques and instruments.
• Successfully generated isotropic vEM volumes from seven public datasets, even without isotropic training data.
• EMDiffuse includes self-assessment for prediction reliability.

Conclusions:

• EMDiffuse significantly enhances the capacity of EM and vEM for large-volume nanoscale imaging.
• The generated isotropic volumes facilitate accurate 3D ultrastructure analysis.
• EMDiffuse is poised to advance research into intricate subcellular structures within large biological systems.