Scan-Rescan reproducibility of carotid bifurcation geometry from routine contrast-enhanced MR angiography

Abstract

Purpose: To demonstrate the feasibility of rapid and reliable geometric characterization of normal carotid bifurcation geometry from routine 3D contrast-enhanced magnetic resonance (MR) angiograms.

Materials and methods: Repeat scans of 61 participants, acquired as part of the Atherosclerosis Risk in Communities (ARIC) Carotid MRI substudy, were digitally segmented using automated 3D level set methods, relying on an operator only to select the branch endpoints and thresholds for the 3D lumen surface initialization. Geometric factors characterizing the 3D lumen geometry were then extracted automatically.

Results: Of 122 scans, 117 could be segmented within 5 minutes each, with 40% being of sufficiently high quality to require less than 2 minutes each. Irrespective of scan quality, geometric factors were found to be highly reproducible, with intraclass correlation coefficients (ICCs) typically above 0.9. The reconstructed lumen surfaces were reproducible to <0.3 mm on average, comparable to previous MRI-based reproducibility studies. Owing to the automated nature of the analysis, operator reliability was near-perfect (ICC >0.99), with lumen surface differences <0.1 mm.

Conclusion: The 3D geometry of the carotid bifurcation can be characterized rapidly and with a high degree of consistency, even for suboptimal image qualities. This bodes well for large-scale retrospective or prospective studies aimed at teasing out the influence of local vs. systemic risk factors for early atherosclerosis.

Figure 1

Key steps in the lumen segmentation process. A: Coronal maximum intensity projection of a representative medium-quality 3D CEMRA dataset. B: Cropped view of VMTK interactive window, showing initialization of the left carotid bifurcation, in this case requiring two stages for the ICA-CCA tract. C: Refined surface after level set evolution, triangulation and surface remeshing.

Figure 2

Representative scan-rescan lumen surfaces and derived planes used for calculating lumen areas. Each panel shows the complete lumen surface segmented from the baseline scan and the lumen surface segmented from the repeat scan after registration and clipping at the CCA3, ICA5 and ECA1 locations. Black contours identify the various planes from the baseline surface, and the number at the bottom right is the RMS difference, in mm, between baseline and lumen surfaces. The left, middle and right 2x2 panels show cases having high, medium, and low pair quality scores, respectively, with cases chosen to span the range of RMS surface differences for each quality group.

Figure 3

Bland-Altman plots for geometric factors. Horizontal and vertical axes are the pairwise mean and difference, respectively. Solid lines identify the mean difference, and dotted lines indicate the 95% confidence limits of agreement, defined as the mean ± RMS difference. Triangles (red), circles (yellow) and squares (green) identify data from high, medium and low quality pairs, respectively.

Figure 4

Histogram of RMS surface differences by pair surface quality.