Segmentation of perivascular spaces in 7T MR image using auto-context model with orientation-normalized features

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Quantitative study of perivascular spaces (PVSs) in brain magnetic resonance (MR) images is vital for understanding the brain lymphatic system and its link to neurological diseases.
• Accurate segmentation of thin, tubular PVS structures in 3D MR images presents a significant challenge due to their complex morphology and orientation.

Purpose of the Study:

• To develop and evaluate a novel learning-based method for accurate segmentation of perivascular spaces (PVSs) in 3D brain MR images.
• To address the challenges associated with segmenting thin, tubular structures with varying directions.

Main Methods:

• A learning-based approach was proposed, involving region of interest (ROI) determination using anatomical structure and vesselness information.
• Randomized Haar features, normalized by principal directions, were extracted within the ROI and classified using a random forest model.
• A sequential learning strategy was employed to incorporate contextual patterns for enhanced segmentation accuracy.

Main Results:

• The proposed method was evaluated on 7T brain MR images from healthy subjects, assessing voxel-wise and cluster-wise accuracy, and geometric properties (volume, length, diameter).
• Performance was also validated on simulated images with motion artifacts and lacunes, demonstrating potential for clinical populations.
• Experimental results confirmed that the proposed method outperformed all existing PVS segmentation techniques.

Conclusions:

• The developed learning-based method provides a robust and accurate solution for perivascular space segmentation in brain MR images.
• This advancement holds significant potential for improving the diagnosis and understanding of neurological diseases associated with PVS abnormalities.