Deep feature regression (DFR) for 3D vessel segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Accurate structural information of coronary arteries is crucial for diagnosing and treating coronary artery disease.
• Existing 3D vessel segmentation methods face challenges in robustness and accuracy.

Purpose of the Study:

• To propose a deep feature regression (DFR) method for robust and accurate 3D vessel segmentation.
• To enhance the quantitative diagnosis and treatment of coronary artery disease through improved vessel segmentation.

Main Methods:

• A convolutional regression network (CRN) was developed for deep feature learning from 600,000 sample images.
• A stable point clustering mechanism was introduced to evaluate estimation reliability and eliminate outliers, increasing tracking robustness.
• A vessel segmentation algorithm was designed using trained deviation parameter estimators with termination criteria based on stable points and intensity constraints.

Main Results:

• The DFR method achieved an average overlapping ratio of 97.5% and an average error of 0.27 mm on a coronary artery dataset.
• Quantitative tests on a cerebral artery dataset showed high accuracy in vessel centerline tracking, with an average error below 0.33 mm.

Conclusions:

• The proposed DFR method demonstrates high accuracy and robustness for 3D vessel segmentation.
• This technique offers significant potential for improving the quantitative diagnosis and treatment planning of cardiovascular diseases.