Self-supervised segmentation and characterization of fiber bundles in anatomic tracing data

Area of Science:

• Neuroscience
• Computational Biology
• Medical Imaging

Background:

• Anatomic tracing is crucial for mapping brain connections and validating diffusion MRI tractography.
• Manual charting of fiber trajectories is time-consuming, limiting available annotated data for validation.
• Developing computer-assisted segmentation methods is essential to accelerate tracer data analysis.

Approach:

• Proposed a novel deep-learning method with a self-supervised loss function for fiber bundle segmentation in macaque brain histological sections.
• Employed semi-supervised learning to leverage unlabeled data, addressing the scarcity of manual labels.
• Incorporated anatomic and across-section continuity constraints to enhance segmentation accuracy.

Key Points:

• The method achieves a true positive rate of approximately 0.80, successfully segmenting unseen sections from different cases after training on a single case.
• Demonstrated the method's utility in quantifying fiber bundle density across different white-matter pathways.
• Revealed varying densities of fiber bundles originating from the same injection site, impacting microstructure-informed tractography.

Conclusions:

• The developed deep-learning approach significantly accelerates the analysis of anatomic tracing data.
• The method's accuracy and ability to generalize across cases offer a valuable tool for neuroscience research.
• Findings on fiber bundle density have implications for refining diffusion MRI tractography and understanding brain circuitry.