Tricuspid valve flow measurement using a deep learning framework for automated valve-tracking 2D phase-contrast

Abstract

Purpose:

Tricuspid valve flow velocities are challenging to measure with cardiovascular magnetic resonance (CMR) since the rapidly moving valvular plane prohibits direct flow evaluation, but they are vitally important to diastolic function evaluation. We developed an automated valve-tracking 2D method for measuring flow through the dynamic tricuspid valve.

Methods:

Nine healthy subjects and two patients were imaged. The approach uses a previously trained deep learning network, TVnet, to automatically track the tricuspid valve plane from long-axis cine images. Subsequently, the tracking information is used to acquire 2D phase-contrast (PC) with a dynamic (moving) acquisition plane that tracks the valve. Direct diastolic net flows evaluated from the dynamic PC sequence were compared to flows from 2D PC scans acquired in a static slice localized at the end-systolic valve position, and also ventricular stroke volumes (SV) using both planimetry and 2D PC of the great vessels.

Results:

The mean tricuspid valve systolic excursion was 17.8±2.5mm. 2D valve-tracking PC net diastolic flow showed excellent correlation with SV by right ventricular planimetry (bias±1.96SD=−0.2±10.4 ml, ICC=0.92) and aortic PC (−1.0±13.8 ml, ICC=0.87). In comparison, static tricuspid valve 2D-PC also showed a strong correlation but had greater bias (p=0.01) vs. RV SV (10.6±16.1 ml, ICC=0.61). In most (8/9) healthy subjects, trace regurgitation was measured at begin-systole. In one patient, valve-tracking PC displayed a high velocity jet (380 cm/s), with maximal velocity agreeing with echocardiography.

Conclusion:

Automated valve-tracking 2D-PC is a feasible route towards evaluation of tricuspid regurgitant velocities, potentially solving a major clinical challenge.

Introduction

Tricuspid valvular regurgitation (retrograde flow of blood from the right ventricle (RV) to the atrium) during systole is often due to elevated RV pressure, as observed in patients with pulmonary hypertension (1) or diastolic dysfunction accompanying heart-failure with preserved ejection fraction (EF). According to the current ACC/AHA guidelines, tricuspid valve regurgitation is assessed with comprehensive transthoracic echocardiography (TTE) with Doppler interrogation of blood velocities (2).

Cardiovascular magnetic resonance (CMR) also provides an effective method for regurgitant volume estimation using the “indirect” planimetry method. This consists in, e.g., for tricuspid regurgitation (TR), subtracting the RV stroke volume (SV) by planimetry from the net pulmonic (or aortic) flow. In recent studies, the value of this CMR planimetry method over TTE has been demonstrated for the mitral valve (3,4); it might similarly have advantages for the tricuspid valve and is considered the reference standard (5).

The indirect CMR approach itself is based on the assumption that only one valve is leaky, and also requires many breath-holds (a cine stack and phase-contrast (PC) acquisition), and depends on accurate flows of the great vessels (6). More importantly, it does not permit diastolic function evaluation to identify elevated LV filling pressures, which requires maximal TR velocity evaluation among other measurements (and not regurgitant volume). A direct evaluation of flow through the valve is challenging because of the large and complex motion of the valve plane. Direct valvular flow measurement would allow estimation of the velocity of the tricuspid regurgitant jet, which is used to derive the pressure gradient and thus RV systolic pressure. Not only does this pressure estimation play a key role in diastolic dysfunction evaluation (7), but it is also needed for initial diagnosis of pulmonary hypertension (PH) (7–9). Such measurements are currently only feasible with TTE, or potentially 4D Flow CMR (10–16). 4D flow CMR has been investigated to quantify mitral and tricuspid regurgitant volume and visualize jets. However, 4D Flow CMR is not used clinically for this purpose and has long scan times and requires trade-offs in temporal and spatial resolution. This means that TTE alone can quantify TR jets (17–19).

Kozerke et al. (20,21) first showed that 2D PC of the valves was possible, using a dynamic slice with valve-tracking, where motion was estimated using a slice tagging method. More recently, semi-automated feature-tracking of the valve plane on long-axis cine images has been developed (22). Based on this more advanced valve-tracking approach, a 2D valve-tracking PC method was tested to directly measure mitral flow and regurgitation (23). This method permitted more practical and accurate tracking compared to earlier slice-following PC approaches.

Here we seek to automate and extend the valve-tracking PC method to the tricuspid valve, (which is characterized by greater and more complex motion), using TVnet. TVnet (24,25) is a previously developed deep learning method for automated measurement of the displacement of these tricuspid valve, including translations and rotations, which processes orthogonal long-axis cine views of the RV (2ch and 4ch), in a fully automatic approach. This information is provided to a 2D PC sequence, modified to incorporate the phasic rotations and translations of the tricuspid valve into the slice prescription during the acquisition.

Materials and Methods.

CMR acquisition

Nine healthy volunteers (36±16y, BMI of 24.9±3.8, 4 females) underwent a CMR exam (3T PRISMA, Siemens Healthcare, Erlangen, Germany). Additionally, two patients were scanned on a 1.5T Siemens (34 and 41 year old males with congenital heart disease) using a 330 and 400 cm/s VENC, to capture any fast regurgitant jets. The study was approved by our local IRB and all subjects provided written informed consent.

The tricuspid valve tracking acquisition workflow is shown in Figure 1, in which first long-axis cines are acquired, exported and processed with the TVnet framework, to obtain tracking information, which is ported to the sequence via a file. The entire semi-automated workflow (mainly saving and loading cine data) requires ~5 minutes (<1minute for TVnet). The long axis cine series included the standard 4-chamber cine and the much less common RV 2-chamber (26), both needed to determine the dynamic tricuspid valve slice acquisition plane. The cine acquisitions had the following average scan parameters: 8–10 second breath-hold, bSSFP, acquisition matrix=157×157, GRAPPA factor 2, repetition time=2.6 ms, echo time=1.29 ms, flip angle=35° voxel size=1.88*1.88*6–8 mm³, 30 frames, 14 views per segment, and 36 ms temporal resolution.

Figure 1:

Tricuspid valve-tracking workflow. Showing (numbered) (1) acquisition of cine RV 2ch and 4ch, (2) automated valve-tracking analysis offline. (3) Meanwhile, other acquisitions can be acquired such as Aortic and PA flow, and a short-axis cine stack, for comparisons of SV. (4) Tracking information is uploaded to the MRI sequence via a file (USB device), followed by (5) valve-tracking PC acquisition. (6) To obtain blood velocity measurements relative to the valve, the valve plane velocity is subtracted.

Automated tricuspid valve tracking (TVnet)

These cine images were exported offline and the tricuspid valve was tracked with TVnet (24,25), as described below. TVnet is a previously developed dual-stage deep learning pipeline, employed to automatically derive tricuspid valve slice geometry. TVnet used transfer learning and is based on ResNet-50, and was trained on 140 and 91 subjects (mostly patients) for 4 chamber and RV 2 chamber cines, respectively, all with 30 time frames expertly annotated. The first stage comprised of a coarse identification of tricuspid valve insertion points, which information was sufficient to guide the standardization of the images in terms of cropping, sizing and orientation, as a preprocessing step; the second stage was trained on standardized images and provided annotation of valve points with high accuracy. The two long axis cine views were processed by two independently trained networks. TVnet tracking network was implemented in MATLAB R2019b (Mathworks, Natick, MA), and is available on GitHub (https://github.com/ra-gonzales/Tvnet). TVnet uses DICOM coordinates (corner position, and readout and phase-encoding vectors) to define the geometry of the long-axis cines. TVnet tracks the valvular insertion points and outputs the 3D coordinates of the tricuspid valve insertion points from both cine views (2ch and 4ch, totaling two line segments, and 4 (x,y,z) coordinates). The data from the two views is combined, so that the center of the tricuspid valve was determined as the mean location of the four insertion points and the valve plane normal $(\overrightarrow{n_t}$, where t indexes the frame) was obtained as the vector product from the two line segments that define the valve plane on the 4ch and 2ch views. This valve plane information, the center point, and normal vector at each time point (for all 30 cine frames interpolated to 500 frames) was saved as a text file for use by the pulse sequence (see Availability of Data statement for code). Processing was automated, except saving and loading cine images. This slice tracking information was automatically saved to a file, during the TVnet processing and automatically retrieved by the valve-tracking PC sequence.

Dynamic tricuspid valve flow sequence

The valve-tracking PC sequence was customized to support dynamic slice prescription (see Video S1) which changes with cardiac phase. The acquisition slice was initialized by the user, centered and oriented according to the tricuspid valve plane at its early-systolic level, using the RV 2- and 4-chamber views. The sequence then automatically retrieves and converts each of the 500 (arbitrarily chosen number of frames for fine interpolation) slice plane positions (i.e. 3 coordinates defining the center-point of the slice, and 3 coordinates defining the normal vector of the slice, $\overrightarrow{n_t}$) needed for direct use by the sequence. The normal to the valve $\overrightarrow{n_t}$ was directly used as the valve-tracking acquisition plane normal. The new image in-plane directions, i.e. phase encode $\overrightarrow{{pe}_t}$ direction and read out vectors $\overrightarrow{{ro}_t}$ direction, were implemented through cross products with $\overrightarrow{n_t}$. The slice prescription was changed during the RR interval, by applying changes to the slice center-point and orientation matrix based on TVnet valve-tracking results, relative to the first frame, rather than an absolute slice prescription.

Scan parameters were: 2D GRE PC, GRAPPA x2, partial Fourier 6/8, acquisition matrix=256×192, repetition time=5.3 ms, echo time=3.4 ms, flip angle=15°, 380×285 mm FOV, voxel size=1.5×2.0×5–6 mm³ for a VENC of 100 cm/s or 150 cm/s through plane for a temporal resolution of 42ms. This required a 20 heart-beat breath-hold.

Immediately following this scan, a standard 2D PC scan with a static slice prescription was acquired, coinciding with the tricuspid valve at its end-systolic level, and with identical scan parameters to the valve-tracking PC acquisition.

Reconstruction and binning

Images from these sequences (both valve-tracking and static) were reconstructed offline using MATLAB R2019b (Mathworks, Natick, MA). The reconstruction used GRAPPA (27) to reconstruct skipped phase-encodings, followed by POCS (28) to synthesize missing k-space, due to partial Fourier acquisition (see raw k-space is Figure S1). As retrospective gating bins data over multiple heartbeats, RR variations can alter the number of cardiac phases acquired for each phase-encoding segment (Figure S1A). With the vendor reconstruction each segment is interpolated to the same number of cardiac phases, which may introduce possible slice-location mismatch of certain phase-encodings and cardiac phases. To limit this, our binning approach assumed that RR variations derived from variations in diastolic but not systolic duration. The temporal binning was thus designed to interpolate (using MATLAB’s “p-chip” method) each heartbeat to the complete number of frames, only from the last ~30% of the RR interval instead of the entire cardiac cycle. This strongly decreased artifacts due to RR variations compared to the scanner reconstruction, as illustrated in Figure S1.

Short-axis cine, and aortic and pulmonary artery flow

For comparison of SVs, a cine short-axis stack (with parameters similar to long-axis cines) and standard 2D PC of the PA and aorta were obtained. The PC acquisitions of PA and aorta used the following parameters: breath-holding, retrospective ECG-gating, acquisition matrix=256*192, TR=5.0 ms, echo time=3.0 ms, flip angle=20°, 4 views per segment, temporal resolution=40 ms, voxel size=1.39*1.39*5–6 mm for a VENC of 150 cm/s through plane. The aortic 2D PC was averaged three times in free breathing while PA flow was performed in a breath-hold, based on their sensitivity to respiration.

Image analysis

Image analyses were performed using Segment software 3.2 R8836 (Medviso, Lund, Sweden) (29). RV and LV volume evaluation was performed on the short-axis stack using the planimetric method with manual contouring of the endocardium in both end-systole and end-diastole. Trabeculations were included as blood pool and the basal slice was determined by reviewing the slice prescription on the long-axis views in both end-systole and end-diastole. Measurements followed consensus recommendations (30). Ventricular SV’s were assessed as the difference of LV or RV volumes between end-diastolic and end-systolic times. Aorta and PA net flow, equivalent to the SV, were assessed using semi-automatic contouring of the vessel wall; background phase correction was performed with static tissue detection, using linear correction.

For both static and dynamic tricuspid PC, the tricuspid valve contours were drawn semi-automatically on the images. Phase correction included spatially varying and cardiac phase dependent background phase correction (first or second order polynomials dependent on the subject) and velocity correction (varied linearly in space) (Figure S2). The latter was applied to correct for blood velocities, so that they were measured relative to the tricuspid valve motion, by subtracting the moving tricuspid plane velocity from the measured blood velocities, as described by Kayser et al. (31). This is necessary to quantify the flow through a moving orifice, and was originally used for static PC. Because there were both translations and rotations of the valve plane, this correction was done pixelwise as shown in Figure S2. Diastolic net flow volume through the tricuspid valve was measured and compared to PA and aorta net flow as well as RV and LV SV. Indirect regurgitant volume (difference between PA or aorta net flow and RV or LV SV, respectively) was compared to the net systolic regurgitant volume from the tricuspid valve 2D PC acquisitions.

The resulting success of the valve-tracking PC acquisition described above, using automated valve-tracking in providing flow at the true valve-plane, was also analyzed. The accuracy of valve-tracking PC magnitude images was analyzed and scored subjectively on a 4-point scale: (1) scan plane did not track tricuspid valve, even in systole; (2) tracking succeeded until systole, but failed in the remainder of the RR interval; (3) tracking succeeded throughout the RR interval, except in a few frames; (4) the tricuspid valve plane was visible throughout the entire RR interval.

Statistical analysis.

All continuous variables are presented as mean ± standard deviation. The different SV were compared using a non-parametric Wilcoxon rank-sum test, Bland and Altman analysis was performed (mean ± 1.96⋅standard deviation of the differences) and both the coefficient of variation (CV) and intra-class correlation (ICC) were reported. All statistical analysis were performed using JMP-SAS software 15, except ICC which were calculated in MATLAB.

Results

Tricuspid valve tracking

TVnet successfully tracked the valve plane on the long axis cine images in 8 of 11 subjects without intervention, and was inadequate in 2 healthy subjects and 1 patient. For those subjects, semi-automatic feature-tracking was used, requiring less than 5 minutes of additional processing time (32,33). Upon later close inspection, for one healthy subject, the cine exhibited some artifacts due to imperfect triggering and tracking was adequate but needed adjustments; in another healthy subject, quality was excellent but tracking showed errors for both RV 2ch and 4ch; in a patient with congenital heart disease, we hypothesize that the unusual anatomy might present a challenge for TVnet.

Furthermore, we evaluated the final success of valve-tracking work-flow (using automated methods, with manual intervention in 3 cases), based on visual inspection of the PC magnitude images of the valve tracking sequence. Almost all studies exhibited successful tracking (n= 9, score of 4), but one study had a score of 2 (only tracking until end-systole), and one had a score of 3 (tracking successful except in a few frames); the mean score was 3.7.

RV 2-chamber and 4-chamber cine images revealed a mean maximal excursion of the tricuspid valve by 17.8±2.5 mm; the maximal valve plane azimuth and elevation angles (averaged over the cohort) were 17±5° and 9±3°, respectively, for healthy subjects. This excursion and angulation occurred during systole and was reversed in diastole.

Tricuspid valve phase contrast

Figure 2 illustrates the static and valve-tracking magnitude and phase images with the resulting flow waveforms in a healthy subject. While the static acquisition, planned at end-systolic valve level, depicted the RV cavity during much of the cardiac cycle, the valve-tracking PC followed the tricuspid valve throughout the cardiac cycle. Note that the valve-tracking PC images closely match the static PC images at end-systole. The flow curves show that relative velocity correction has an important effect on flow values. This can be observed in systole when the tricuspid valve is closed and the correction causes expected near-zero flows, for both static and dynamic planes (net systolic flow was −1.1±4.3 ml and −2±6.4 ml for the static and valve-tracking respectively), where negative flow volumes indicate slight regurgitation.

Figure 2:

Valvular total flow in a healthy control. A) Flow images (magnitude and velocity maps) for static and dynamic slices. Top: Static slice. Bottom: Valve-tracking slice, with the RV valve visible at each point in the cardiac cycle. ROIs for flow quantification are shown in two phases, with an attempt to include only valvular region in the ROIs. B) Flow curves for static (blue) and dynamic (red) planes showing an E and A wave in diastole, and minimal flow in systole. Note that correction for valve velocity (solid lines) yeilds physiologically expected lower flows in systole, vs. uncorrected (dashed lines). Diastolic and systolic intervals are indicated, and the corresponding time-points on the flow curves matching each imaging time in A) are marked (black dots). The increased peak flow in early diastole using the static vs. dynamic approach is evident (black arrow); this lead to static PC providing higher SVs than valve-tracking and overestimating forward flow (vs. RVSV) in general. Note that valve-tracking reduces the (non-physiological) measurement of antegrade flow during systole (green arrow) vs. static. Using valve-tracking, there is visualization of some regurgitation at end diastole during valve closure (orange arrow) not detected by static PC.

At end-diastole, some regurgitant flow is measured by the valve-tracking method, which cannot be observed in the static slice, in this example. Most (8/9) of our healthy subjects presented with similar short-lived regurgitant flow during the first or last frames of the cardiac cycle that might result from a delay between RV contraction and the closing of the tricuspid valve. However, these were not “jets” since regurgitant flow was observed over the entire valve. In this subject, the flow volume for the static slice is larger vs. valve-tracking (see arrow at 400 ms). This exemplifies the general finding that using the static end-systolic slice over-estimated diastolic flow.

Mean SV, aortic and PA flow volumes, and direct tricuspid valve flow volumes are presented Table 1. Without correction of the phase-velocities for the relative motion of the valve, the valvular flows underestimated the SV (i.e. were lower than SV by Ao and PA flow, and by planimetry). The valve-plane velocity correction improved both static and valve-tracking phase-contrast measurements, disclosing strong agreement with reference SV measures, and reducing systolic flow measurements towards zero, which is expected without regurgitation. Table 2 provides the Bland-Altman analysis, correlation and ICC between the reference SV or great vessels net flow and the diastolic net flow from tricuspid valve PC –both static and dynamic. As expected in healthy subjects, there were excellent correlations between RV and LV SV and net flows through the aorta (R>.97), but less strong for the PA (R>.79). Tricuspid valve net flow in diastole was compared against each of these measures, from both static and dynamic scan planes. There were strong correlations (R>.79, excluding the PA). However, the static slice presented significantly higher bias (p=0.01 by paired t-test) than the dynamic slice. The ICC for agreement of valvular PC with RVSV was higher with the dynamic slice (0.92 vs. 0.61) and the CoV was smaller (4.6% vs. 11.3%), as displayed in Figure 3. Table S1 shows the same data but calculated for net flow (instead of diastolic flow only), in which apparent flow, even during systole when the valve was closed, was included in the measurements. While the Pearson rho coefficients of valvular PC vs. RVSV are higher for static PC, vs. valve-tracking (0.89 vs. 0.83), the ICCs show that valve-tracking PC is more accurate (0.66 for static vs. 0.82), as do the CoVs.

Table 1.

Ventricular planimetry stroke volumes and 2D PC flow of the great vessels and tricuspid valve

	Total flow		Systolic flow	Diastolic flow
LVSV [ml]	79 ± 12		-	-
RVSV [ml]	77 ± 11		-	-
AoSV [ml]	78 ± 12		-	-
PASV [ml]	74 ± 11		-	-
Static tricuspid valve net flow [ml]
Uncorrected	87 ± 12		16 ± 5	73 ± 12
Corrected	87 ± 12		−1 ± 5	88± 15
VT tricuspid valve flow [ml]			(net)	(net)
Uncorrected	74 ± 11		12.6 ± 9.8	62 ± 10
Corrected	75 ± 12		−2.0 ± 6.4	77 ± 14
Regurgitation	[ml]	[%]
RVSV - PASV	3.6 ± 6.8	4.4
RVSV - AoSS	−0.8 ± 3.3	−0.8
RVSV - LVSV	−2.0 ± 2.5	−2.5

Ao: aorta, LV/RV: left/right ventricle, PA: pulmonary artery, SV: Stroke Volume

Table 2.

Analysis of agreement between measurements of stroke volumes or net diastolic flow

		Stroke volume or Diastolic net flow
		Limits of Agreement [ml]	Pearson Rho (p-value)	ICC (95% CI)	CoV [%]
RVSV vs.
	LVSV	2 ± 4.8	0.98 (>.0001)	0.96 (0.77, 0.99)	2.8
	AoSV	0.8 ± 6.4	0.97 (>.0001)	0.96 (0.85, 0.99)	2.9
	PASV	−3.6 ± 13.3	0.81 (0.0077)	0.79 (0.34, 0.95)	6.9
	Static TV flow *	10.6 ± 16.1	0.84 (0.0044)	0.61 (−0.1, 0.9)	11.3
	Tracked TV flow *	−0.2 ± 10.4	0.94 (0.0002)	0.92 (0.67, 0.98)	4.6
LVSV vs.
	AoSV	−1.2 ± 4.4	0.99 (>.0001)	0.98 (0.91, 1)	2.2
	PASV	−5.6 ± 12.4	0.85 (0.00383)	0.77 (0.15, 0.95)	7.5
	Static TV flow	8.7 ± 17.7	0.79 (0.01052)	0.65 (−0.02, 0.91)	10.3
	Tracked TV flow	−2.2 ± 12.7	0.89 (0.00147)	0.87 (0.56, 0.97)	5.9
AoSV vs.
	PASV	−4.4 ± 15.1	0.79 (0.01049)	0.76 (0.27, 0.94)	7.9
	Static TV flow	9.9 ± 17.3	0.81 (0.0086)	0.64 (−0.06, 0.91)	9.1
	Tracked TV flow	−1 ± 13.8	0.86 (0.00288)	0.87 (0.52, 0.97)	7.1
PASV vs.
	Static TV flow	14.2 ± 22.8	0.64 (0.06446)	0.4 (−0.13, 0.81)	15.7
	Tracked TV flow	3.4 ± 19.9	0.69 (0.03857)	0.68 (0.12, 0.92)	9.5

Bland-Altman limits of agreement, Pearson Rho and coefficient of variation (CoV) between the reference stroke volumes or great vessels flow and tricuspid valve (TV) flow.

Ao: aorta, LV/RV: left/right ventricle, PA: pulmonary artery, SV: Stroke Volume;

^*p=0.01 by paired t-test.

Figure 3:

Linear regression plots of RV Stroke volume (RVSV) compared to diastolic net flow volume in healthy subjects by A) static PC with slice plane coinciding with the valve plane at end-systole, and B) valve-tracking PC, prescribed at the valve plane. The flow values were corrected for both eddy currents and for relative valvular velocities. Both methods showed excellent correlation to RVSV, but valve-tracking PC flow volumes exhibited lower bias (p=0.01 by paired t-test) and coefficient of variability, and a slope closer to unity. This was reflected by an excellent ICC for valve-tracking PC vs. RVSV (ICC=0.92 for valve-tracking PC vs. 0.61 for static PC).

The indirect method for calculating regurgitant flow showed very low extent of regurgitation, so that the accuracy of direct regurgitant flow could not be evaluated in these healthy subjects, as expected.

Tricuspid valve E and A waves during diastole

RV filling E and A wave peak velocities were measured from valve-tracking and static flow. Mean E, A and E/A using valve-tracking PC were 45±4, 33±3 cm/s, 1.4±0.3, and mean E, A and E/A using static PC was 48±8 cm/s, 33±5 cm/s, 1.5±0.3, respectively. These were in accordance with normal TTE values from the literature (34).

Tricuspid diastolic peaks were similar between both static and dynamic acquisition. Although E/A was well correlated for static and dynamic PC (r=0.83, p=0.01), we found that E (r=0.67, p=0.07) and especially A (r=0.11, p = 0.79) were not.

Case studies

Figure 4 shows data from a 34 year old man with with D-transposition of great arteries status post arterial switch for the indirect approach measured 23 ml regurgitation. Dynamic and static PC found similar regurgitant volumes (26 and 16 ml), where quantifications are shown in Figure 4.

Figure 4:

Tricuspid flow in patient with congenital heart disease, with mild tricuspid regurgitation, comparing static and valve-tracking flow (both planned just below the valve at begin-systole). A) The dynamic slice plane permitted imaging the valve throughout the cardiac cycle, as observed on magnitude images. Note that the static slice shows the right atrial blood pool in end-systole, while the valve-tracking flow maintains the valvular position. B) Tricuspid flow wave-forms for static and dynamic planes show similar curves. The indirect approach yielded a RVSV of 135 ml, net forward flow of 112 ml (Aortic net flow), and regurgitant volume of 23 ml. Using valve tracking, the forward flow volume was 129 ml, net forward flow was 103 ml, and the regurgitant volume was 26 ml, while for static PC, these values were 129, 113, and 16 ml. C) Flow maps show that the regurgitation is somewhat more persistent using dynamic PC than static PC. The VENC for valve-tracking PC was 330 cm/s (since a high velocity jet was anticipated), vs. 150 cm/s for the static acquisition, leading to noisier flow maps for dynamic PC. E) TTE shows a jet in early systole, with a peak jet velocity of ~150 cm/s.

The valve-tracking images showed that even at end-systole, the imaging plane maintained its location on the valve (Figure 4A), unlike the static flow images. The flow maps show some valvular regurgitation appears to persist somewhat longer during systole with valve-tracking (Figure 4C) than with the static slice; this is an expected key advantage of our technique. With a static slice, the regurgitant jet moves easily out of plane. The TTE continuous wave doppler is also shown, with a maximal jet velocity of 150 cm/s, which was not observed by 2D PC (maximal velocity of ~60 cm/s for both static and dynamic PC). This may be due to the slice prescription, which was chosen directly at the valve plane; the vena contracta (the location of highest velocity) is known to be downstream of the regurgitant valve.

Figure 5 shows results from valve-tracking PC in a 41 year old man with congenital heart disease. The jet is visible on the long-axis cine (Figure 5A). The valve-tracking PC successfully captured the same tricuspid valve plane throughout systole (Figure 5AB, chosen slightly superior to the valve plane), and showed a high velocity jet on the phase image (inset, VENC =400 cm/s). Compared to a recent TTE, the jet had a similar maximal velocity (Figure 5C). The notch in the TR jet by CMR might be physiological or might be due to an artifact, e.g. loss of the slice position (but that is not evident). This patient did not have complementary sequences to measure regurgitant volume via the indirect approach.

Figure 5:

Tricuspid flow in a patient with congenital heart disease. A) The jet is visible on the long-axis cine. The valve-tracking plane was prescribed above the valve. B) The valve tracking permitted precisely imaging the plane just above the valve throughout systole. C) The peak velocities by valve-tracking PC and TTE agreed well (both about 380 cm/s).

Discussion.

In this study, we demonstrated the usefulness of combining a previously trained deep learning framework (24) (25) for valve-tracking, and the technique of valve-tracking phase-contrast (20,21), which had not been applied to the tricuspid valve. We modified a 2D-PC sequence to rotate and translate following the tricuspid valve trajectory during the cardiac cycle. This allowed a direct evaluation of the tricuspid valve flow in one breath hold using an automated analysis for dynamic slice planning, which was shown to be more accurate than a static slice. Indeed, we found excellent agreements between net flow using valve-tracking PC when compared to net flows in the great vessels and RV and LV planimetry, in healthy subjects. Finally, we showed a case example, where a high tricuspid regurgitant jet velocity was recorded by CMR, and compared well with TTE.

In most of the subjects, the fully automated valve-tracking method based on a deep learning network, TVnet, was used successfully. The tricuspid valve anatomy was clearly visualized throughout the cardiac cycle. In three cases, the automated tracking required manual intervention. TVnet can thus be further improved by additional training data, e.g. more patients and including data with lower quality, and unusual anatomy.

The obtained tricuspid flow curves exhibited an expected physiological pattern with classic diastolic E and A waves matching well with TTE literature velocities values (34), especially when using valve-tracking. We often observed a transient backward flow in healthy subjects, which can be likened to regurgitation, over the full valve area, either in begin-systole or late-diastole in the healthy subjects, but have not found prior studies reporting this. We hypothesize that this happens just before the closing of the valve, when some blood volume may be pushed back into the right atrium by the leaflet during the tricuspid valve closing.

The patient case studies show the promise of this technique; since valve-tracking flow revealed regurgitant flow that persisted longer in systole compared to a static slice, which is unable to fully assess the regurgitation as the valve moves in systole. Also, the regurgitant volumes estimated using the dynamic 2D-PC sequence was comparable to the indirect method combining RV planimetry and great vessels net flow. Finally, we were able to accurately measure peak velocity of the tricuspid regurgitant jet, with agreement to TTE, in a patient with a very rapid jet velocity (380 cm/s). To our knowledge, a quantitative example where CMR (e.g. by 4D Flow MRI) captured a high velocity TR jet has not been presented.

Other solutions for measuring flow through the moving valve have been proposed. 4D flow CMR has been explored to quantitatively assess tricuspid valve and mitral valve flow (12,13) by retrospectively tracking the valves. One 4D Flow CMR study used retrospective tracking of the tricuspid valve along the cardiac cycle from two orthogonal planes (10). It found that a static 2D-PC acquisition showed significant overestimation of tricuspid flow, while underestimating the regurgitant fraction (35). However, 4D Flow CMR has limitations on temporal and spatial resolution (36,37) and acquisition times are long (~5–10 minutes). We were not able to compare the valve-tracking 2D PC with 4D flow measurements (using retrospective valve-tracking). Driessen et al. (10) compared 4D forward flow of the tricuspid valve (with valve-tracking and valve velocity correction) vs. PA flow, which yielded Bland Altman limits of agreement of −1.6 +/−18.5ml, while 2D static PC overestimated forward flow, similar to the findings of our study.

Another approach used two 2D PC acquisition plane along the valve trajectory and select the plane closest to the valve in each frame for post-processing for a retrospective tracking(38). This involves acquisition and processing of additional scans increasing scan time, and complexity. Merging information from the multiple PC scans might also be complex and was not evaluated in the prior study, which instead compared the flow values from the different slice planes, demonstrating different and complementary information. They found that the optimal static slice prescription for measuring RV filling (RVSV) coincided with the end-systolic valve plane, while a begin-systolic plane was better for measuring regurgitant flow. Therefore, the choice of a slice plane coinciding with the valve at end-systole for static PC was appropriate in this study of mainly healthy subjects and measurement of forward tricuspid flow. However, a slice plane planned to coincide with the valve (or above the valve) at end-diastole is the right choice for capturing regurgitation (for both static and dynamic PC).

The tricuspid valve has not been studied with prior valve-tracking 2D PC methods. Rotations were not included in a recent valve tracking study of the mitral valve (23), but seemed to have been incorporated in earlier valve-tracking (20,21). Rotations are highly challenging to implement, but might be needed for the tricuspid valve. The angulation measured in subjects (~17° and 9°) between an acquisition plane and the main flow direction (if uncorrected) could lead to a velocity underestimation. The blood velocity data was corrected for relative motion of the valve, including a complicated spatially varying correction to account for angular velocity. It might be that a simpler velocity correction scheme would work more robustly.

This deep learning valve-tracking approach is versatile, and would also work for the mitral valve, for which a similar deep learning based valve-tracking method has already been developed (MVnet, trained with extensive multicenter data) (39,40).

Limitations

This study included mainly healthy subjects without tricuspid regurgitation. Further study in a large patient cohort is highly warranted. The PA flow appeared less strongly correlated to other SVs. This may be due to the influence of cardiac motion, or simply to less consistent planning of this acquisition plane. Minor artifacts were generated using the dynamic slice, but were almost eliminated using our off-line reconstruction, with careful binning of retrospectively acquired k-space. TVnet, the deep-learning network, can be improved with more training data sets to increase its success rate. Finally, the accuracy of tricuspid flow was validated using surrogates of RVSV and LVSV, and aortic and pulmonary artery flow, which are imperfect gold standards.

Conclusion.

In conclusion, we have developed a dynamic PC sequence with fast and automated prospective tracking of the tricuspid valve for direct flow evaluation within a breath. This could lead to faster and more accurate (compared to the indirect method) evaluation of tricuspid regurgitant volume, as well as regurgitant jet velocities.

Supplementary Material

Supinfo

Figure S1: Binning of retrospectively ECG-gated k-space data. A) K-space acquired with retrospective ECG-gated phase-contrast, showing k-space signal in the phase-encoding vs. cardiac frame domain. The central k-space is fully sampled for most of the cycle. Partial Fourier in the phase-encoding direction is evident from the missing region of k-space. GRAPPA is evident in that there is R=2 undersampling of non-central k-space. The last few frames show missing k-space data due to variable RR interval. The k-space data is temporally interpolated and the 2D PC data is reconstructed into magnitude and velocity images, with results shown in B) and C) for the last frames (end-diastole). B) The vendor-provided interpolation appears to interpolate over the full cardiac cycle, while our customized interpolation C) interpolates only the final 30% of the RR (end-diastolic k-space data). End-diastolic only interpolation strongly reduced ghosting artifacts (C), that were present in the vendor-reconstruction and are exacerbated by the dynamic slice. There is also a timing difference observed (arrows on phase-image) between the vendor and custom reconstruction. In our experience, comparing retrospectively and prospectively ECG-gated PC (without binning), the timing is more accurate (similar to non-binned data) using diastolic-only binning.

Figure S2: Phase-contrast corrections included time-dependent background phase corrections based on static tissue identification, as shown in (A), performed in Segment, and (B) corrections based on velocity of the mitral valve (which varied spatially due to rotations). The measured valve velocity at each pixel within the valve was subtracted from the blood velocities. (C) Phase correction provided in cm/s from the background phase correction (top) and from the displacement correction (bottom). D) The mean valve velocity (averaged over the valve) for each subject, and the mean velocity averaged over all volunteers (black line).

Table S1: Analysis of agreement between total tricuspid flow (including systole and diastole) and stroke volumes.

Movie S1. Video S1: (TR_valveplane.mp4).

Illustration of a dynamic slice, which follows the tricuspid valve plane throughout the cardiac cycle, planned from the RV 2ch and 4ch, generating dynamic phase-contrast.

List of Abbreviations

TR

tricuspid regurgitation

CMR

cardiovascular magnetic resonance

PC

phase-contrast

TEE

transesophageal echocardiography

PA

pulmonary artery

SV

stroke volume

RV

right ventricle

LV

left ventricle

Availability of data and materials:

The sequence is available customer to customer via Siemens, and the valve-tracking methodology TVnet is available on GitHub. The imaging data can be provided upon reasonable request. The Matlab code for generating this dynamic slice plane prescription based on long-axis cines is provided on GitHub (https://github.com/dcpeters777/Valvetracking).