Abstract
Quantitative assessment of coronary artery stenosis from 2D X-ray angiograms is often limited by qualitative clinical evaluation and significant image noise. This paper presents an automated pipeline for the extraction and graph-based analysis of the coronary artery tree. By integrating morphological Linear Top-Hat (LZH) operators with modified Gabor filters (MGF), the method segments both primary and peripheral vessels. The resulting skeleton is converted into a mathematical graph to facilitate objective pre- and post-operative diagnostics. Evaluation against expert ground truth demonstrates a correct detection rate of 72.41%, though challenges remain regarding false positives from medical instrumentation and image artifacts.
Introduction
Diagnosing and treating coronary artery stenosis relies heavily on X-ray angiograms. Currently, clinicians often evaluate the success of procedures, such as angioplasty, through qualitative visual inspection of blood flow regeneration in finer vessels. Automating this process is technically challenging because the heart's constant motion precludes standard background-removal techniques like Digital Subtraction Angiography (DSA). Consequently, researchers must extract complex, tree-like structures from single, static images cluttered with irregular backgrounds including lungs, ribs, and radiation noise.
Methodology
The authors propose a three-step pipeline to transform raw angiograms into analyzable graph structures:
- Vessel Segmentation: An "artificial" DSA approach is employed, combining a Linear Top-Hat (LZH) Operator—a morphological filter using line-shaped elements to isolate vessels—and a Modified Gabor Filter (MGF). The MGF is run in two passes to capture both thick main arteries and fine peripheral vessels. The results are fused by taking the maximum pixel values across filtered images.
- Skeletonization: The extracted grayscale vessels are binarized and thinned to 1-pixel-wide lines, preserving the topology of the vessel tree.
- Graph Analysis: The skeleton is converted into a mathematical graph. Morphological masks are used to identify nodes (bifurcations and crossings) and edges (the vessels themselves), enabling quantitative flow analysis.
Results and Evaluation
The system was evaluated against a ground truth of 10 manually skeletonized angiograms using three spatial metrics (with a 2-pixel tolerance):
- $F_1$ (Correctly Detected): The system achieved an average correct detection rate of 72.41%.
- $F_2$ (False Positives): The system detected non-existent vessels 34.29% of the time, often misidentifying catheters or image borders.
- $F_3$ (False Negatives): The system missed existing vessels 19.83% of the time, primarily in high-noise regions.
Qualitatively, the system successfully maps the general coronary tree. However, it struggles with "hallucinated" connections between parallel vessels and the inherent ambiguity of 2D projections, where vessel crossings can be mistaken for true bifurcations.
Discussion and Future Work
While the hybrid filtering approach effectively captures both major and minor vessels, the high false positive rate remains a significant limitation. Future work will focus on implementing sub-pixel accurate skeletonization (e.g., Steger’s algorithm) to reduce spatial deviations. Additionally, more sophisticated topological algorithms are needed to differentiate between vessel crossings and bifurcations, eventually providing clinicians with a quantitative "plausibility measure" to support diagnostic confidence.
Summary updated