Scalable Biologically-Aware Skeleton Generation for Connectomic Volumes.

Area of Science:

• Neuroscience
• Computational Biology
• Data Science

Background:

• Connectomic datasets are growing rapidly, necessitating efficient methods for neuronal skeleton generation.
• Existing skeletonization techniques often lack biological accuracy or computational efficiency.

Purpose of the Study:

• To develop a novel topological thinning strategy for accurate and efficient neuronal skeleton generation.
• To incorporate biological constraints into the skeletonization process for improved accuracy and relevance.
• To enhance the analysis, evaluation, and visualization of large-scale connectomic data.

Main Methods:

• Implemented a biologically-informed topological thinning strategy for centerline generation.
• Pre-processed segmentation data to remove biologically-infeasible bubbles, improving accuracy and speed.
• Utilized a Convolutional Neural Network (CNN) to detect cell bodies for skeleton anchoring.
• Employed a synapse-aware topological thinning approach for detailed skeleton reconstruction.
• Estimated neurite width and geodesic distance between synapses and cell bodies.

Main Results:

• Achieved significant improvements in accuracy (47.5% for width, 62.8% for distance) over baseline methods.
• Demonstrated processing speeds exceeding one million voxels per second per CPU.
• Showcased linear scalability for large-scale connectomic datasets.
• Validated results on over 1250 neurons across three species.

Conclusions:

• The proposed method generates accurate, biologically relevant neuronal skeletons efficiently.
• This approach significantly advances the computational pipeline for connectomics research.
• The method offers a scalable and robust solution for analyzing massive neuronal datasets.