Robust Deep 3D Blood Vessel Segmentation Using Structural Priors.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning significantly enhances 3D blood vessel segmentation accuracy.
• Challenges persist with limited labeled training data and noise robustness.
• 3D vessel structures exhibit informative 2D edge profiles on image slices.

Purpose of the Study:

• Propose a novel deep 3D vessel segmentation network guided by edge profiles.
• Address limitations of scarce training data and ensure noise robustness.
• Improve accuracy and efficiency in 3D blood vessel segmentation.

Main Methods:

• Developed a deep network with a shared encoder and two decoders for joint segmentation and edge profile learning.
• Integrated bidirectional convolutional long-short term memory (BCLSTM) modules for 3D context.
• Introduced novel regularization terms for local homogeneity and noise robustness.

Main Results:

• The proposed network outperforms state-of-the-art methods on benchmark datasets using DICE and mean Intersection-over-Union.
• Performance gains are substantial with limited training data.
• Achieved low computational cost during network inference compared to existing methods.

Conclusions:

• The edge-profile-guided deep learning network effectively improves 3D blood vessel segmentation.
• The method demonstrates robustness to noise and efficiency, especially in data-scarce scenarios.
• This approach offers a promising solution for accurate and efficient 3D vessel segmentation in medical imaging.