Topology aware multitask cascaded U-Net for cerebrovascular segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Accurate cerebrovascular segmentation is vital for diagnosing cerebrovascular diseases.
• Traditional deep learning methods struggle with the complex topology of blood vessels.
• Existing differentiable skeletonization methods are often inaccurate or slow.

Purpose of the Study:

• To develop a more accurate and efficient method for cerebrovascular segmentation.
• To improve the topological representation of segmented cerebrovascular networks.
• To address limitations of current differentiable skeletonization techniques.

Main Methods:

• A cascaded multitask U-Net architecture was proposed to directly compute vascular skeletons.
• The model was trained using the clDice loss for topological constraint embedding.
• Magnetic resonance angiography (MRA) images were used for training and evaluation.

Main Results:

• The proposed U-Net achieved accurate cerebrovascular segmentation with improved topology.
• The method demonstrated a good balance between accuracy and computation time.
• It outperformed current state-of-the-art methods in topological accuracy.

Conclusions:

• The cascaded U-Net offers a robust and efficient solution for cerebrovascular segmentation.
• Directly learning vascular skeletons improves topological accuracy in segmentation.
• This approach facilitates better computer-aided diagnosis for cerebrovascular pathologies.