RH vorticity at peak diastole was compared between normal subjects and subjects with PAH and RVDD. Same-day 4D Flow and echocardiographic data was acquired for each subject. 4D Flow images were preprocessed to improve data quality, and qualitative and quantitative characteristics were assessed using ParaView visualization and quantification software (Kitware, Clifton Park, NY, USA) [21
2.1. Data Acquisition
The study population consisted of 20 subjects with RVDD and 14 controls. All subjects gave their informed consent for inclusion before they participated in the study. The study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by the Institutional Review Board of National Jewish Health (Protocol #2808. Approved 1/23/14). RVDD subjects had RVDD per American Society of Echocardiography guidelines [22
]. Echo-derived diastolic function was used as a surrogate for invasive RV end diastolic pressure measurements which may lead to over or under estimation of RV diastolic impairment. In order to exclude conditions that may confound RH flow patterns, subjects (both control and RVDD) with cardiomyopathy, coronary artery disease, significant valvular heart disease, or advanced liver disease were excluded from analysis.
Two-dimensional and Doppler echocardiography were performed to obtain LV and RV diastolic parameters including MV and TV early (E) and late (A) filling peak velocities, E/A ratio, early diastolic deceleration time, and lateral and septal tricuspid and mitral early (e′) and late (a′) diastolic velocities. RVDD was defined as either stage I (TV E/A < 0.8, TV E/e′ > 6, and DT > 120 ms) or stage II (TV E/A = 0.8–2.1, E/e′ > 6, and DT > 120 ms) [23
Time-resolved 3D phase-contrast cardiac MRI imaging was performed using a Siemens Avanto 1.5 T MRI scanner with the subject in the supine position. In-plane voxel dimensions were square and ranged from 1.98 mm to 2.60 mm depending on patient size, with a slice thickness of 3 mm for all patients. The field of view was rectangular with voxel volumes ranging from 11.75 to 20.35 mm3
. Other scan parameters were α = 15°, TE/TR = 2.85/48.56 ms, venc = 100–150 cm/s and temporal resolution was 50 ms. An RF-spoiled gradient echo pulse sequence, prospective ECG gating, and respiratory navigators were used as described in [24
]. Note that because of multi-cardiac cycle time averaging of the velocity data that occurs with MRI scans, stochastic properties are not present in the final velocity data which instead represents an ensemble average over multiple cardiac cycles.
In additional to phase-contrast images, steady state free precessing axial, short axis, 4-chamber, and 2-chamber 2D cine images were obtained for morphological and functional assessment. Short axis images with in-plane resolutions ranging from 1.09 to 1.56 mm, slice thickness of 6 mm, and voxel volumes ranging from 7.2 to 14.6 mm3 were obtained from the ventricular apex to beyond the tricuspid plane.
Several 4D Flow data preprocessing tools and techniques were used in this study that have significant impact on the resulting data quality, including noise reduction and anti-aliasing algorithms.
The noise reduction algorithm was based on the tissue magnitude image method of Bock et al. [25
] in which velocity data is set to zero in regions with low signal-to-noise. The tissue-contrast intensity threshold used for this study was set at 14 for all images which corresponds to approximately 3% of the maximum pixel intensity for the images. Velocity was set to zero in voxels corresponding to an intensity value of 14 or less in the spatially and temporally corresponding anatomical magnitude image. Visual analysis of noise-filtered images for a range of thresholds and subjects showed this value to be a good compromise between noise reduction and unwanted removal of velocity data.
Venc was initially set to 100 cm/s for an initial round of subjects due to measurements of bulk velocities that were not predicted to exceed this value. However, after noting significant aliasing in these scans, particularly in the MPA and ascending aorta during peak systole and in the region of the TV in patients with TV regurgitation, venc was increased to 150 cm/s for the remainder of the subjects, although some aliasing still occurred. To correct velocity aliasing, an iterative anti-aliasing algorithm was developed similar to that of Axel and Morton [26
]. Shown in Figure 1
is a logic chart for the anti-aliasing algorithm, and shown in Figure 2
is a representative result from several iterations of the algorithm on a heavily aliased image. The code is set to terminate after a maximum of 10 iterations.
2.3. Vorticity Calculation
Vorticity was calculated in ParaView using a first order bilinear interpolation scheme over the entire RH for all subjects at each time-step in the cardiac cycle. The volume integrated vorticity workflow involved preprocessing raw data as previously described, converting velocity data and cine images to Ensight and VTK formats respectively, visualizing velocity and vorticity vectors in ParaView, thresholding vorticity vectors, summing vorticity vector magnitudes within a rectangular prismatic region of interest (containing either RA, RV, or RA + RV volumes), and multiplying the resulting summation by voxel volume to get a volume integration.
The combined RA + RV volume was manually isolated using a rectangular prism oriented along the RA/RV longitudinal axis using early diastolic vorticity vectors and 2D short-axis and 4-chamber MRI images as a visual guide (see Figure 3
). The rectangular prism was oriented such that RH diastolic vorticity vectors were included in the volume, while any vorticity in the LV and ascending aorta was excluded. A subjective judgment was often made regarding how much of the inferior and superior vena cava to include in the RA—which may have an effect on the results, particularly in patients with diastolic backflow in the vena cava for which significant diastolic vorticity was present in the veins. However, an interobserver variability study, described elsewhere in this paper, indicates that the effect is small. In order to separate the RV from the RA, the early diastolic 4-chamber tissue contrast image was used to assist in locating the tricuspid plane at peak diastole.
2.4. Estimation of Cardiac Event Timing
The diastolic phase of the cardiac cycle is split into an early diastolic phase (E-wave) during which blood is drawn into the ventricle via ventricular relaxation, and late diastole (A-wave) in which atrial contraction forces additional blood into the ventricle. The timing of peak E-wave in the RH—defined here as the moment of greatest time rate of ventricular volume change during early diastole—was estimated using the peak of the tricuspid flow rate time series (described below). When a distinct E-wave peak was less apparent (generally in RVDD subjects), curves of the time derivative of LV volume were used in addition to TV flowrate to estimate peak early diastolic timing in the RV. The length of time between E- and A-waves was determined from the LV volume curves and was then subtracted from the A-wave TV flow peak time to yield the RV E-wave time for vorticity analysis. In addition, flowrates through the MPA were calculated, which, when combined with the TV flowrates, were used for comparison between right and left heart events.
Flowrates through the TV and MPA were calculated using 4D Flow velocity data and ParaView flow visualization software. The TV plane perpendicular to the direction of bulk blood flow was initially located visually using a cine animation of a mid-TV 4-chamber MRI tissue contrast slice at the approximate start of diastole as shown in Figure 4
a. Figure 4
b shows a spherical region of interest (“clip” in ParaView) combined with an interpolated slice of blood velocity data co-located with the TV plane to produce a disk of velocity data approximating the area of flow through the TV. The size and location of the disk was further refined by visually comparing it to early diastolic velocity magnitude images in the TV plane, the full field of 3D velocity vectors, and high-resolution short-axis tissue contrast images during all early diastolic time-steps (Figure 4
c,d). The area-integrated normal flow through the disk was then calculated for each time-step using the ParaView Surface Flow filter and the resulting TV flowrate time-series was exported from ParaView. Due to movement of the tricuspid annulus during diastole, this method is only accurate at the time-step for which the disk position is optimized.
A plane approximately perpendicular to the MPA was located for each subject in a region between the pulmonary valve and the left/right pulmonary artery split using axial cine images. An initial location was found in which spatial separation between the MPA and aorta during systole was large enough to prevent aortic flow from contaminating the MPA velocity data. A method similar to that described above for TV flow was then used to refine the size and location of a disk over which normal velocity was integrated to produce a time series of MPA flowrate (See Figure 5
). Because spatial mismatch can occur between 4D Flow and cine data, final location of the disks for TV and MPA flowrate calculation was performed using 4D Flow velocity data rather than cine data.
The LV volume as a function of time was determined from short-axis cine data. Ideally, RV volume would be used, but due to the complexity of RH geometry, automatic segmentation schemes are still in development and require high computational times relative to automatic LV segmentation [27
]. The LV endocardium of each subject was semi-automatically segmented at each short-axis cine image time-step using the research version of Medviso’s Segment software, v1.9 R3763 (Medviso AB, Lund, Scania, Sweden) [28
]. Each segmentation boundary curve was visually inspected for accuracy, and in rare cases in which the curve differed considerably from the visible endocardium, the curves were manually corrected in Segment. Segmentation curves were converted to LV volume curves within Segment and exported. The resulting LV volume time-series were exported for each subject and the central difference time derivative of volume for each subject was calculated at all time-steps. Forward difference and backward difference was used for the first and last time-steps respectively.
Isolating the RA, RV, or RA + RV volumes using a rectangular prism permits the undesirable influence of contributions from small regions of velocity that may lie outside the physical boundaries of the RH chambers. These velocities may be due to blood vessels, velocity noise in tissue regions, or tissue motion. Both these velocities and associated vorticity levels were low, so in order to reduce the sensitivity of the results on the subjective placement of the rectangular prismatic region of interest, small vorticity vectors were removed from the region of interest (ROI) prior to volume integration by setting all vorticity magnitudes below a threshold value to zero. This resulted in integration of the large vorticity vectors only, and a vorticity-free buffer region near the prismatic surface walls. Thus, small changes in surface placement have minimal effect on the integration, and numerical problems with partial cell integrations are avoided. Vorticity thresholding involved deciding on a vorticity magnitude threshold level for a single healthy subject, and then scaling that value for the remainder of the subjects using a cardiac flow parameter. Cardiac index (CI) was ultimately used as a vorticity threshold scaling parameter due to its incorporation of several parameters including heart rate, stroke volume, and body surface area. By scaling the threshold value to CI, the effect of body size and the flowrate of the heart are accounted for, resulting in a larger inter-subject effect due to changes in flow structures. In addition to the aforementioned CI scaled thresholding, vorticity was also thresholded at constant levels across all subjects in order to determine the impact of the choice of scaling parameter.
The resulting thresholded vorticity magnitude element summations were multiplied by voxel volume to get the numerical spatial integration of vorticity. Integrated vorticity was then scaled by CI to again reduce the dependency of the results on body mass, heart size, and heart rate; factors that are expected to influence overall vorticity. The resulting scaled and spatially integrated vorticity was then used as our metric of interest.
RH vorticity at peak E-wave was calculated by two trained analysts using the method previously described for an initial 23-subject cohort consisting of 13 RVDD patients and 10 controls. The concordance correlation coefficient was calculated between the two analysts’ results for constant (unscaled) threshold levels ranging from 0.005 to 0.100 s−1
in 0.005 s−1
increments in order to examine the effect of threshold level on interobserver reliability. A flow chart of postprocessing steps is shown in Figure 6