1349444 - Automated Cardiac Valve Tracking and Segmentation for Assessment of Valvular Blood Flow using 4D flow MRI

Cardiac valvular flow quantification from 4D flow MRI is a valuable diagnostic method for assessing valve function and intracardiac hemodynamics, but analysis involves time-consuming manual valvular segmentation. The purpose was to compare 4D flow MRI with automated valve tracking and novel automated segmentation to manual valve segmentation in patients with corrected atrioventricular septum defect (AVSD) and healthy volunteers.

Methods:

Data was retrospectively collected from 8 patients with corrected AVSD (mean age, 20 ± 8 years) and 23 healthy volunteers (mean age, 26 ± 14 years). Free breathing whole-heart 4D flow MRI was acquired at 3T without respiratory gating. After automatic valve tracking time-resolved trans-valvular velocity images were reformatted on which cardiac valve annuli were segmented using CAAS MR Solutions (Pie Medical Imaging). Automatic segmentation was performed using a deformable model which integrated local phase coherency. Fully manual and fully automated segmentation was performed by one observer and manually corrected automated segmentation by two observers. Net forward volumes (NFVs) and regurgitation fractions (RFs) were calculated for all valves. Analysis times, NFVs and RFs were compared by Wilcoxon signed-rank test. NFV variation was calculated as the mean NFV of all valves over the standard deviation of differences among all four valves. Intra- and interobserver variability was tested by intraclass correlation coefficients (ICCs).

Results:

Interobserver agreement for fully automated versus fully manual segmentation was excellent for NFV (ICC ≥ 0.97) and strong for left and right atrioventricular valve (LAVV, RAVV) RF (ICC = 0.89 and 0.92, respectively). Intraobserver agreement for manually corrected automated segmentation was excellent for NFV (ICC ≥ 0.99) and LAVV and RAVV RF (ICC ≥ 0.96). No differences in NFV variation over cardiac valves was observed for fully automated versus fully manual segmentation (10.4% [IQR, 7.3%-16.0%] vs 9.6% [IQR, 6.9%-12.9%], respectively; P = 0.164). LAVV and RAVV RF were significantly smaller for fully automated versus fully manual segmentation (LAVV, 3.1% [IQR, 2.0%-4.3%] vs 4.1% [IQR, 2.8%-5.1%], respectively; P = 0.001. RAVV, 7.1% [IQR, 4.4%-10.6%] vs 8.1% [IQR, 6.4%-10.9%], respectively; P = 0.009). Analysis time for all subjects was shorter for manually corrected automated versus fully manual segmentation (4.1 minutes [IQR, 2.9-4.9 minutes] vs 9.4 minutes [IQR, 7.8-10.5 minutes], respectively; P < 0.001). Analysis time for manually corrected automated segmentation was small but significantly different between observers (4.1 minutes [IQR, 2.9-4.9 minutes] vs 4.8 minutes [IQR, 3.6-6.7 minutes], P = 0.006).

Conclusion:

Manually corrected automated segmentation does not compromise quantification of NFV and LAVV and RAVV RF in patients with corrected AVSD and healthy volunteers. Valvular flow quantification with 4D flow MRI benefits from automated valve segmentation by reducing post-processing analysis time.