A novel end-to-end deep learning solution for coronary artery segmentation from CCTA

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Diseases

Background:

• Coronary computed tomographic angiography (CCTA) is crucial for diagnosing cardiovascular diseases.
• Automatic coronary artery segmentation (CAS) is a challenging but vital task in CCTA analysis.
• Accurate CAS aids in precise diagnosis and treatment planning for cardiovascular conditions.

Purpose of the Study:

• To propose a novel, end-to-end deep learning (DL) solution for automated coronary artery segmentation (CAS).
• To enhance the accuracy and efficiency of CAS in CCTA images.
• To develop a DL model that preserves vessel integrity, including shape details and continuity.

Main Methods:

• A fully automatic, multistage DL solution inspired by the Di-Vnet network was developed.
• The solution was trained and validated on 338 CCTA cases with pre-annotated cardiac and coronary artery masks.
• Performance was evaluated using Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (95% HD), Recall, and Precision.

Main Results:

• The proposed DL solution achieved high accuracy: 90.29% DSC, 2.11 mm 95% HD, 97.02% Recall, and 92.17% Precision.
• The method demonstrated significant efficiency, processing images at 0.112 s/image and cases at 30 s/case on average.
• Performance surpassed existing state-of-the-art segmentation methods.

Conclusions:

• The novel DL solution effectively automates CAS in an end-to-end manner.
• The method achieves a simultaneous high level of accuracy, efficiency, and robustness.
• This automated approach holds promise for improving cardiovascular diagnostics through CCTA analysis.