A Simple Method for Automatic 3D Reconstruction of Coronary Arteries From X-Ray Angiography

Area of Science:

• Cardiovascular Imaging
• Medical Artificial Intelligence
• Biomedical Engineering

Background:

• Three-dimensional (3-D) reconstruction of coronary arteries (CA) from medical imaging, particularly X-ray coronary angiography (XCA), remains a complex challenge.
• Accurate identification of vessel ends is crucial for precise 3-D CA reconstruction and subsequent functional analysis.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for automatic identification of the proximal and distal ends of coronary arteries in XCA.
• To assess the efficacy of using averaged template models of CA for matching 2-D segmented vessels from different XCA views.
• To compare the accuracy and clinical relevance of the proposed deep learning and template model methods against manual matching.

Main Methods:

• A U-net deep learning architecture, incorporating a Resnet encoder and decoder, was trained on manually labeled XCA images to identify vessel ends.
• Training datasets comprised 2,342 (LAD), 1,907 (LCX), and 1,523 (RCA) labeled images.
• Averaged 3-D template models from ten reconstructions were used for matching 2-D segmented vessels from two XCA angles, with comparisons to manual matching.

Main Results:

• The deep learning network successfully identified the proximal region of LAD, LCX, and RCA in 97.7%, 97.5%, and 96.4% of test images, respectively.
• Success rates for identifying the distal region were 94.9% (LAD), 89.8% (LCX), and 94.6% (RCA).
• No statistically significant differences were found in projection distances or computed fractional flow reserve (FFR) between template model matching and manual matching.

Conclusions:

• Deep learning is a feasible and accurate method for automatically identifying coronary artery ends in XCA.
• Template model-based matching offers accuracy comparable to manual correspondence for 3-D coronary artery reconstruction from bi-plane XCA.
• The developed methods facilitate more efficient and potentially more accurate 3-D coronary artery reconstruction.