Detecting coronary artery abnormalities is crucial in the management of coronary artery disease. Traditionally CT angiography (CTA) is the modality of choice to detect stenosis, however, acquisition and image quality of coronary MR angiography (CMRA) has improved considerably over the past years^1,2,3. Image-based respiratory motion correction yields high-quality images by free-breathing, and allows accurate depiction of the coronary arteries and lumen narrowing stenosis⁴. For time-efficient analysis automatic centerline tracking is required. This has extensively been investigated in CTA⁵, but only before recent improvements in acquisition were made in CMRA⁶. The aim of this study was to perform automatic coronary artery centerline tracking on state-of-the-art CMRA data, in a clinically relevant population.

Data

Thirty patients with suspected coronary artery disease were included. CMRA images (acquired resolution 1.3 mm isotropic, reconstructed ~0.74x0.74x0.65 mm) were acquired as described in⁴. A 2D image navigator was generated by spatially encoding the startup echoes, and gating was achieved using a diminishing variance algorithm⁷. Clinical records indicated that 9 patients had single or multi-vessel disease based on previous CTA or X-ray imaging.

Centreline tracking

For reference manual centerlines were annotated on the three main coronary artery branches: right coronary artery (RCA), left anterior descending (LAD) and left circumflex (LCX). Occluded (3) and stented (2) arteries were excluded.

Automatic centerline extraction consisted of two steps: 1) vesselness computation and 2) fast marching. The vesselness filter⁸ has widely been used for vascular analysis in medical images. This filter enhances vessel-like structures by using the eigenvalues of the Hessian matrix, which is composed by local second order derivatives of the image⁸. The implementation of the vesselness filter involves several parameters that need to be optimized based on the target vessel and image type. Based on the expected size of coronary arteries we computed the Hessian matrix at three scales: 0.5, 1 and 1.5 mm, and for each voxel took the maximum vesselness response. The parameters α and β were used to balance between tube- and plane-like structures and the deviation from blob-like structures, and were optimized on a random selection of three patients (9 arteries). Combinations for α and β of 0.3, 0.5, 0.7 and 0.9 were evaluated. Centrelines were traced between the start and end point of each manually annotated centerline. The fast marching algorithm was used to compute a vesselness-based distance map with respect to the start of the centerline. Subsequently the shortest path from the end point to this start point was traced. The settings that yielded the smallest average centerline distance between the automatic and manual centerlines were used to analyse the arteries of the remaining 27 patients. All implementations were done in Matlab.

Evaluation

Results were visually assessed for correct tracing, and if errors were detected a second automatic tracing was performed using one additional point on the manual centerline in the area where the tracing went wrong. Secondly, centerline distances were evaluated. Both manual and automatic centerlines were resampled to a 0.1mm spacing. For each point on each automatic centerline the shortest distance to the manual centerline was determined. Lastly, the length of each manual centerline was determined, as well as the distance of this centerline that could be traced correctly automatically.

The optimized parameters for vesselness computation were α=0.9 and β=0.3. Four examples of obtained vesselness images are shown in Figure 1. All (quantitative) results are provided in Figure 2. Examples of automatically traced centerlines are shown in Figures 3 and 4.Automatic centerline extraction from CMRA is possible in 74% of arteries in clinically relevant patients, increasing to 82% when one additional point on the centerline could be used. Lowest accuracy was found for LCX, which was mainly caused by wrong tracing through neighboring veins. However, generally, these results show promise for automatic coronary artery analysis on CMRA.Automatic centerline extraction is possible from CMRA data. Future work will focus on automated stenosis detection.