A high-performance deep-learning-based pipeline for whole-brain vasculature segmentation at the capillary resolution.

Area of Science:

• Neuroscience
• Medical Imaging
• Computational Biology

Background:

• Accurate whole-brain vascular mapping is crucial for understanding brain function and disease mechanisms.
• High-resolution whole-brain imaging generates massive datasets, posing challenges for efficient vessel segmentation.
• Vessel segmentation is a fundamental step in reconstructing and analyzing the entire brain vasculature.

Purpose of the Study:

• To develop a high-performance, deep learning-based pipeline for rapid and accurate whole-brain vessel segmentation.
• To design a lightweight deep neural network capable of extracting multi-resolution vessel features for precise segmentation.
• To validate the pipeline's efficiency and accuracy on large-scale whole-brain vascular data.

Main Methods:

• Introduced HP-VSP, a parallelized deep learning pipeline comprising data blocking, block prediction, and block fusion.
• Developed a lightweight deep neural network utilizing multi-resolution feature extraction for scalable vessel segmentation.
• Validated the approach using whole-brain vascular data from three transgenic mice acquired via HD-fMOST imaging.

Main Results:

• The HP-VSP pipeline achieved state-of-the-art performance in whole-brain vessel segmentation across various metrics.
• The segmentation network demonstrated high accuracy with significantly reduced parameters (1% of similar networks).
• The pipeline successfully completed whole-brain vessel segmentation in 3 hours and is applicable to vascular analysis.

Conclusions:

• HP-VSP provides an efficient and accurate solution for whole-brain vessel segmentation from large-scale imaging data.
• The developed lightweight deep neural network offers a computationally efficient approach to complex vascular feature extraction.
• This pipeline facilitates advancements in whole-brain vascular atlas construction and disease mechanism studies.