vEMINR: Ultra-Fast Isotropic Reconstruction for Volume Electron Microscopy With Implicit Neural Representation

Area of Science:

• Neuroscience
• Biophysics
• Computational Biology

Background:

• Volume electron microscopy (vEM) provides nanoscale 3D visualization of biological structures.
• vEM suffers from anisotropic resolution, with significantly lower axial than lateral resolution due to sectioning limitations.

Purpose of the Study:

• To introduce vEMINR, an ultra-fast implicit neural representation (INR) based method for isotropic vEM image reconstruction.
• To enhance vEM image quality by learning true degradation patterns and accelerate reconstruction using INR's efficient parameterization.

Main Methods:

• Developed vEMINR, an implicit neural representation (INR) method for vEM image reconstruction.
• Utilized INR's continuous function representation and efficient parameterization for accelerated processing.
• Trained the model to learn true degradation patterns from low-resolution vEM images.

Main Results:

• vEMINR achieved over tenfold faster reconstruction compared to mainstream methods across 11 public datasets.
• Demonstrated significantly higher accuracy in reconstructing organelles and neurons from vEM data.
• Maintained reconstruction accuracy while enabling high-throughput processing of terabyte-scale vEM datasets.

Conclusions:

• vEMINR offers a substantial improvement in reconstruction speed and accuracy for vEM datasets.
• The method's efficiency facilitates large-scale vEM image analysis and related research.
• vEMINR is poised to become a significant tool in advanced biological imaging and research.