TriSAM: Tri-Plane SAM for Zero-Shot Cortical Blood Vessel Segmentation in VEM Images.

Area of Science:

• Neuroimaging
• Microscopy
• Computational Biology

Background:

• Microscale Volume Electron Microscopy (vEM) offers detailed vascular insights but lacks standardized benchmarks for neuroimaging analysis.
• Existing imaging techniques at macro and mesoscales are well-resourced, contrasting with the underdevelopment of microscale vEM benchmarking infrastructure.
• Accurate segmentation of cortical blood vessels in vEM images is crucial for understanding brain structure and function.

Purpose of the Study:

• Introduce the first public benchmark, BvEM, for cortical blood vessel segmentation in vEM images.
• Address the gap in neuroimaging by providing a standardized dataset for evaluating segmentation methods.
• Develop and validate a novel 3D segmentation method, TriSAM, for vEM data.

Main Methods:

• Created the BvEM benchmark using vEM image volumes from adult mouse, macaque, and human brains.
• Standardized resolution and addressed imaging variations, followed by meticulous semi-automatic and manual annotation for high-quality 3D segmentation.
• Developed TriSAM, a zero-shot 3D segmentation method leveraging a multi-seed tracking framework with the SAM model for long-term vessel segmentation without fine-tuning.

Main Results:

• The BvEM benchmark provides high-quality 3D annotated cortical blood vessels across three species.
• TriSAM demonstrated superior performance in zero-shot 3D cortical blood vessel segmentation on the BvEM benchmark.
• The multi-seed tracking framework effectively extends 2D segmentation models to 3D volumes for neuroimaging analysis.

Conclusions:

• The BvEM benchmark is a significant contribution to microscale neuroimaging, facilitating the development and evaluation of vEM analysis tools.
• TriSAM offers a novel and effective approach for 3D blood vessel segmentation in vEM data, outperforming existing methods.
• This work establishes a foundation for advanced microscale vascular analysis in neuroscience research using vEM.