Introduction
Accurate extraction of coronary artery centerlines from Cardiac Computed Tomography Angiography (CTA) is a prerequisite for the quantitative analysis of cardiovascular diseases. Despite its importance, manual extraction is time-consuming and requires significant expertise. Automated "minimal path" algorithms have been proposed but typically suffer from two primary limitations: the requirement for manual seed point initialization and a lack of robustness in complex, noisy anatomical environments where "vesselness" filters may fail due to calcifications or adjacent structures.
This paper presents a fully automatic framework using a Directional Minimal Path (DMP) algorithm. By integrating a prior anatomical model, the system automates initialization and provides directional constraints that guide the tracking process through challenging regions.
Methodology
The DMP framework operates in three distinct stages to ensure both automation and tracking accuracy.
1. Model Establishment
The process begins by aligning a prior anatomical model—containing pre-traced centerlines—to the target patient's CTA volume. This is achieved through a multi-step deformable registration. This step provides the algorithm with initial "seed points" (start points) and a rough estimate of the vessel trajectories, solving the "ostia" initialization problem without manual intervention.
2. Model-Guided Centerline Extraction
The core of the tracking algorithm is a cost function designed to find the optimal path. Unlike traditional methods that rely solely on local vessel-like features, the DMP cost function incorporates three terms:
- Vesselness: Standard tube-detection based on Hessian matrix eigenvalues.
- Intensity Similarity: Ensuring the path stays within blood-filled lumens.
- Directional Guidance: A prior-driven term that penalizes paths deviating from the direction suggested by the registered model. This prevents the tracker from jumping to nearby veins or taking incorrect turns at arterial junctions.
3. Iterative Model Correction
To account for individual anatomical variations that global registration cannot capture, the system employs iterative model correction. As the centerline is extracted, the prior model is locally re-aligned to the newly found path, ensuring that the directional guidance remains accurate throughout the tracking process.
Experimental Results
The method was evaluated using eight cardiac CTA volumes, generating 80 synthetic cases and 56 real-world "leave-one-out" test cases.
Quantitative Performance
The DMP method demonstrated superior robustness compared to the standard Vesselness-based Minimal Path (VMP) approach. In synthetic tests, DMP achieved a 97.5% detection ratio, while the VMP succeeded in only 50% of cases, even when VMP was provided with manual seed point assistance.
In real-world testing across 56 cases, the mean accuracy was recorded at 1.32 ± 0.81 mm. The system showed high reliability across major branches, with detection rates of 92.9% for the Right Coronary Artery (RCA) and 82.1% for the Left Anterior Descending (LAD) artery. The slightly lower performance in the LAD is attributed to its higher geometric variability compared to the RCA.
Conclusion
The proposed Model-Guided Directional Minimal Path framework successfully automates the extraction of coronary centerlines while significantly improving tracking robustness. By treating vessel extraction as a globally guided navigation problem rather than a purely local search, the system overcomes common pitfalls like branch discrimination and image noise. Future work will focus on expanding the method to extract the full coronary tree and validating the approach on larger clinical datasets.
Summary updated