PMC

Validation of PC‐MR flow measurements in a pulsatile phantom experiment. The plots show modified Bland–Altman analyses comparing timer‐and‐beaker flow measurements with flow from 2D PC‐MR using the proposed semi‐automatic delineation method (top panel) and by manual delineation (bottom panel). For both delineation methods, PC‐MR was in close agreement with timer‐and‐beaker measurements at 1·5T (open triangles) and 3T (open squares).

Improvement in segmentation accuracy using the proposed shape constraints in two example cases. The left image shows a transversal image slice used for flow measurement in the ascending aorta, and the right image shows a double‐oblique image slice used for flow measurements in the pulmonary artery, both in ventricular diastole. Semi‐automatic inaccurate segmentations using an optimized active contour curvature force for shape constraints are shown as solid white lines. Improved semi‐automatic segmentations using the proposed method with shape‐constrained reconstruction are shown as dashed white lines.

Example of a 2D PC‐MR flow volume measurement. Top panel (a) shows reference delineations (dashed white lines) of the ascending aorta in a magnitude image (left) and the corresponding phase‐contrast image (right) in early ventricular systole in a transversal image plane. The lower panel (b) shows measured flow over time after delineations in all time phases throughout the cardiac cycle. The flow volume was calculated from the flow sum over time.

Interobserver variability of flow volumes for the proposed semi‐automatic method (top panel) and manual delineations between two observers (bottom panel). Both panels show Bland–Altman analysis. Dotted lines indicate zero flow volume difference, solid lines indicate bias, and dashed lines indicate 95% limits of agreement (LoA). A clear reduction in interobserver variability of measured flow volumes was observed for the proposed semi‐automatic method compared to manual delineation. AO, ascending aorta (open circles); Pulm, pulmonary artery (open squares). The required time of analysis for an experienced observer was approximately 2 min for manual delineation and approximately 10 s for semi‐automatic delineation.

The semi‐automatic segmentation method agreed with reference delineations for flow volumes in the aorta (top panel; n = 151 subjects), flow volumes in the pulmonary artery (middle panel; n = 40 subjects) and Qp/Qs ratio calculations (bottom panel; n = 25 subjects). Left panels show correlation plots between semi‐automatic and manual measurements. Dotted lines indicate lines of identity, and solid lines indicate linear regressions. Right panels show modified Bland–Altman analysis for semi‐automatic and manual measurements. Dotted lines indicate zero difference between compared methods, solid lines indicate bias, and dashed lines indicate 95% limits of agreement (LoA). Low bias and variability were found for the proposed segmentation method compared to reference delineations for flow volume measurements in both aorta and pulmonary artery and for Qp/Qs ratio. AO, ascending aorta; Pulm, pulmonary artery.

The proposed segmentation method resulted in similar performance when initialized at different time points of the RR interval. Top panel shows flow profiles over the RR interval averaged over all subjects for the ascending aorta (left) and the pulmonary artery (right). Middle panel shows flow volume bias and 95% limits of agreement (LoA; filled circles and error bars) of the semi‐automatic method versus reference delineations. Bottom panel shows average Dice coefficients and 95% limits of agreement (LoA; filled squares and error bars). The proposed segmentation method resulted in similar flow volume bias (filled circles; middle left panel), flow limits of agreement (error bars; middle left panel) and Dice coefficient performance (bottom left panel) when initialized at different time points for the ascending aorta. For the pulmonary artery, however, flow volume bias and limits of agreement were slightly sensitive to the time point of initialization (middle right panel). Pulmonary artery segmentations initialized within 10–55% of the RR interval showed stable flow volume bias (filled circles; middle right panel), flow volume limits of agreement (error bars; middle right panels) and Dice coefficient performance (bottom right panel). AO, ascending aorta; Pulm, pulmonary artery.