Petascale pipeline for precise alignment of images from serial section electron microscopy

Area of Science:

• Neuroscience
• Computer Science
• Image Processing

Background:

• Accurate neural circuit reconstruction relies on precise alignment of serial section electron microscopy (ssEM) images.
• Current alignment methods struggle with artifacts and the petascale data sizes of modern ssEM datasets.
• Improved alignment is crucial for advancing automatic image segmentation and subsequent neural circuit mapping.

Purpose of the Study:

• To develop a robust and scalable computational pipeline for aligning ssEM images.
• To enhance the accuracy of 3D image stack creation for neural circuit reconstruction.
• To address challenges posed by image artifacts and large datasets in ssEM analysis.

Main Methods:

• Utilized self-supervised convolutional neural networks (CNNs) trained with metric learning for initial image pair alignment.
• Implemented iterative fine-tuning for precise alignment refinement.
• Employed vector voting for increased robustness against image artifacts and missing data.
• Developed a block-based processing strategy with distributed computation for scalability.
• Achieved global alignment by composing block transformations with decay, avoiding global optimization.

Main Results:

• Demonstrated improved alignment accuracy on a whole fly brain dataset compared to existing methods.
• Showcased the pipeline's scalability to a cubic millimeter of mouse visual cortex data.
• Validated the robustness of the alignment process in the presence of image artifacts.

Conclusions:

• The developed computational pipeline significantly enhances the accuracy and scalability of ssEM image alignment.
• This advancement facilitates more reliable neural circuit reconstruction from large-scale neuroimaging data.
• The publicly available open-source Python packages enable broader adoption and further research in connectomics.