A U-Net Deep Learning Framework for High Performance Vessel Segmentation in Patients With Cerebrovascular Disease

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate brain vessel segmentation is crucial for cerebrovascular disease management.
• Traditional segmentation methods are labor-intensive and lack robust validation.
• Deep learning offers a potential solution for automated and reliable vessel segmentation.

Purpose of the Study:

• To optimize and evaluate a U-net deep learning framework for brain vessel segmentation.
• To compare the performance of the U-net model against traditional graph-cuts segmentation.
• To assess the clinical applicability of deep learning for cerebrovascular imaging.

Main Methods:

• Utilized a U-net deep learning architecture trained on labeled data from 66 cerebrovascular disease patients.
• Evaluated model performance using Dice coefficient, 95% Hausdorff distance (95HD), and average Hausdorff distance (AVD).
• Compared U-net performance with the graph-cuts segmentation method on 2D patches.

Main Results:

• U-net models achieved high performance (Dice ~0.88, 95HD ~47, AVD ~0.4 voxels).
• Demonstrated excellent segmentation of large vessels and sufficient performance for small vessels.
• U-net significantly outperformed graph-cuts (Dice ~0.76, 95HD ~59, AVD ~1.97).

Conclusions:

• Deep learning U-net models provide a highly effective tool for brain vessel segmentation in cerebrovascular disease.
• The U-net approach surpasses traditional methods, encouraging clinical adoption.
• Future research should address segmentation of small vessels and challenging pathologies.