Here we present a protocol for measuring fetal blood flow rapidly with MRI and retrospectively performing motion correction and cardiac gating.
Magnetic resonance imaging (MRI) is an important tool for the clinical assessment of cardiovascular morphology and heart function. It is also the recognized standard-of-care for blood flow quantification based on phase contrast MRI. While such measurement of blood flow has been possible in adults for decades, methods to extend this capability to fetal blood flow have only recently been developed.
Fetal blood flow quantification in major vessels is important for monitoring fetal pathologies such as congenital heart disease (CHD) and fetal growth restriction (FGR). CHD causes alterations in the cardiac structure and vasculature that change the course of blood in the fetus. In FGR, the path of blood flow is altered through the dilation of shunts such that the oxygenated blood supply to the brain is increased. Blood flow quantification enables assessment of the severity of the fetal pathology, which in turn allows for suitable in utero patient management and planning for postnatal care.
The primary challenges of applying phase contrast MRI to the human fetus include small blood vessel size, high fetal heart rate, potential MRI data corruption due to maternal respiration, unpredictable fetal movements, and lack of conventional cardiac gating methods to synchronize data acquisition. Here, we describe recent technical developments from our lab that have enabled the quantification of fetal blood flow using phase contrast MRI, including advances in accelerated imaging, motion compensation, and cardiac gating.
Comprehensive assessment of the fetal circulation is necessary for monitoring fetal pathologies such as fetal growth restriction (FGR) and congenital heart disease (CHD)1,2,3. In utero, patient management and planning for postnatal care depend on the severity of the fetal pathology4,5,6,7. Feasibility of fetal blood flow quantification with MRI and its applications in assessing fetal pathologies have recently been demonstrated3,8,9. The imaging method, however, faces challenges, such as increased imaging times to achieve high spatiotemporal resolution, lack of cardiac synchronization methods, and unpredictable fetal motion10.
Fetal vasculature comprises small structures (~5 mm diameter for major blood vessels that comprise the descending aorta, ductus arteriosus, ascending aorta, main pulmonary artery, and superior vena cava11,12,13).To resolve these structures and to quantify flow, imaging at high spatial resolution is required. Moreover, the fetal heart rate is about twice that of an adult. A high temporal resolution is thus also required to resolve dynamic cardiac motion and blood flow across the fetal cardiac cycle. Conventional imaging at this high spatiotemporal resolution requires relatively long acquisition times. To address this issue, accelerated fetal MRI14,15,16 has been introduced. Briefly, these acceleration techniques involve undersampling in the frequency domain during data acquisition and retrospective high-fidelity reconstruction using iterative techniques. One such approach is compressed sensing (CS) reconstruction, which allows reconstruction of images from heavily undersampled data when the reconstructed image is sparse in a known domain and undersampling artifacts are incoherent17.
Motion in fetal imaging presents a major challenge. Motion corruption can arise from maternal respiratory motion, maternal bulk motion or gross fetal movement. Maternal respiration leads to periodic translations of the fetus, whereas fetal movements are more complex. Fetal movements can be classified as localized or gross10,18. Localized movements involve motion of only segments of the body. They typically last for about 10-14 s and their frequency increases with gestation (~90 per hour at term)10. These movements generally cause small corruptions and do not affect the imaging area of interest. However, gross fetal movements can lead to severe image corruption with through plane motion components. These movements are whole body movements mediated by the spine and last for 60-90 s.
To avoid artifacts from fetal motion, steps are first taken to minimize maternal motions. Pregnant women are made more relaxed using supportive pillows on the scanner bed and dressed in comfortable gowns and may have their partners present beside the scanner to reduce claustrophobia19,20. To mitigate effects of maternal respiratory motion, studies have performed fetal MR exams under maternal breath-hold21,22,23. However, such acquisitions must be short (~15 s) given the reduced breath-hold tolerance of pregnant subjects. Recently, retrospective motion correction methods have been introduced for fetal MRI14,15,16. These methods track fetal motion using registration toolkits and correct for motion or discard uncorrectable portions of acquired data.
Finally, postnatal cardiac MR images are conventionally acquired using electrocardiogram (ECG) gating to synchronize data acquisition to the cardiac cycle. Without gating, cardiac motion and pulsatile flow from throughout the cardiac cycle are combined, producing artifacts. Unfortunately, the fetal ECG signal suffers from interference from the maternal ECG signal24 and distortions from the magnetic field25. Hence, alternative non-invasive approaches to fetal cardiac gating have been proposed, including self-gating, metric optimized gating (MOG) and doppler ultrasound gating21,26,27,28.
As described in the following sections, our MRI approach to quantify fetal blood flow leverages a novel gating method, MOG, developed in our laboratory and combined with motion correction and iterative reconstruction of accelerated MRI acquisitions. The approach is based on a pipeline in a previously published study14 and is composed of the following five stages: (1) fetal blood flow acquisition, (2) real-time reconstructions, (3) motion correction, (4) cardiac gating, and (5) gated reconstructions.
All MRI scans were performed with informed consent from volunteers as part of a study approved by our institutional research ethics board.
NOTE: The methods described below have been used on a 3T MRI system. The acquisition is performed using a radial phase contrast MRI sequence. This sequence was prepared by modifying the readout trajectory (to achieve a stellate pattern) of the manufacturer's Cartesian phase contrast MRI. The sequence and sample protocols are available upon request through our C2P exchange platform. All reconstructions in this work were performed on a standard desktop computer with the following specifications: 32 GB memory, 3.40 GHz processor with 8 cores, and 2GB graphic card with 1024 compute unified device architecture (CUDA) cores. Image reconstruction was performed on MATLAB. Nonuniform fast Fourier transform (NUFFT)29 was performed on the graphics processing unit (GPU). Motion correction parameters were calculated using elastix30. Figure 1 depicts the protocol in a chronological order, tracking how the acquired velocity encodes (color coded in Figure 1) are processed with representative images at each stage of reconstruction. The reconstruction code is available at https://github.com/datta-g/Fetal_PC_MRI. While we provide the steps in the protocol here, most of these algorithm steps are automated in our pipeline.
1. Subject positioning and localizer exams
2. Acquisition of fetal blood flow data
3. Motion correction of fetal measurements
4. Solving for fetal heart rate
5. Reconstruction of fetal CINEs
In general, phase MRI examinations of flow target six major fetal vessels: the descending aorta, ascending aorta, main pulmonary artery, ductus arteriosus, superior vena cava, and umbilical vein. These vessels are of interest to the clinician as they are often implicated in CHD and FGR, influencing the distribution of blood throughout the fetus9. A typical scan duration with the radial phase contrast MRI is 17 s per vessel such that the scans are short while also allowing time for enough data acquisition for CINE reconstruction. The total acquisition time, including localizers and phase contrast MRI, for the representative results was 3 min. In this study, representative results are presented using flow acquisition data from the descending aorta in two human fetuses: Fetus 1 and Fetus 2 with gestational ages (week + days) of 35+4 and 37+3, respectively.
As in Figure 1, initial real-time reconstructions (temporal resolution: 370 ms) performed for motion tracking took 45 s per reconstructed slice. Translation motion tracking took 2 min for each slice. The extracted motion parameters for Fetus 1 (Figure 2 A1, maximum displacement: 1.6 mm) and Fetus 2 (Figure 2 A2, maximum displacement: 1.3 mm) depict the motion of the descending aorta over the duration of the scan. The shared mutual information of each real-time frame with all other co-registered frames are shown in Figure 2 B1 (Fetus 1) and Figure 2 B2 (Fetus 2). In these cases, all frames shared mutual information above the cut off criteria, so no data was rejected. The second real-time reconstructions (temporal resolution: 46 ms), used to derive cardiac gating information, took 10 min for each slice. MOG derived the fetal heartbeat (RR) intervals using a multiparameter model, as shown in Figure 2 C1 (Fetus 1, RR interval: 521 ± 20 ms) and Figure 2 C2 (Fetus 2, RR interval: 457 ± 9 ms).
Final CINE reconstructions using the retrospectively motion-corrected and gated data took 3 min per slice. The anatomical and velocity reconstructions for Fetus 1 and Fetus 2 at peak systole are shown in Figure 3. Reconstructions with motion correction show vessels with sharper walls. Without motion correction, the descending aorta is blurrier and less conspicuous. The measured flow curves from each fetus (Figure 4) show higher peak and mean flows in the reconstructions without motion correction ([peak mean]: Fetus 1 [25.2 9.8] ml/s, Fetus 2 [34.6 10.3] ml/s]) than in those with motion correction ([peak mean]: Fetus 1 [23.5 9.2] ml/s, Fetus 2 [28.7 9.7] ml/s]).
Figure 1: Pipeline to reconstruct fetal phase contrast MRI data. (A) Step 1: Golden-angle radial phase contrast MRI data (color coded as: flow compensation = red & through plane encode = blue). The alternating colors depict that the flow-compensated and through-plane encoded acquisitions occur at the same spatial frequencies. (B) Step 2: Temporal windows of 370 ms for real-time reconstruction using CS with sparsity constraints (STV and TTV). Motion correction and data rejection are performed. (C) Step 3: Temporal windows of 46 ms are created for real-time reconstruction with CS (with STV and TTV sparsity constraints) for MOG. (D) Step 4: The data is binned into cardiac phases (CP), and CS is used to create a fetal flow CINE, with sparsity constraints (STV and TTV). Representative reconstructions from each CS step are shown in the Reconstructions column. Reconstructions for steps 3 and 4 are shown for a time point corresponding to peak systole. Scale bars in the top left corner of the anatomical images denote 10 mm in the image. The time specifications, in seconds, highlighted in grey represent the durations of the corresponding steps. STV: spatial total variation, TTV: temporal total variation, CS: compressed sensing, MOG: metric optimized gating, CINE: gated dynamic reconstruction. Please click here to view a larger version of this figure.
Figure 2: Representative displacement and heart rate curves. A1 and A2 depict retrospectively tracked displacement curve for the scans in Fetus 1 and Fetus 2, respectively. B1 and B2 show the sum of the mutual information of a given frame with all other frames for Fetus 1 and Fetus 2, respectively. The red dotted lines represent 1.5x interquartile range below which data is rejected. C1 and C2 depict the RR intervals derived with MOG in Fetus 1 and Fetus 2, respectively. RR interval: time between consecutive heartbeats, MOG: metric optimized gating. Please click here to view a larger version of this figure.
Figure 3: Representative velocity sensitive CINE reconstructions at peak systole. Each quadrant depicts the anatomical and velocity reconstructions. The top row shows the CINE with motion correction in Fetus 1 and Fetus 2, respectively. The bottom row shows the CINE without motion correction in Fetus 1 and Fetus 2, respectively. The red and blue arrows depict the descending aorta. Scale bars in the top left corner of the anatomical images denote 10 mm. Please click here to view a larger version of this figure.
Figure 4: Representative flow curves in the fetal descending aorta. The solid and dashed data lines depict the flow curves obtained from CINE reconstructions with and without motion correction, respectively, in Fetus 1 (left) and Fetus 2 (right). Please click here to view a larger version of this figure.
This method enables the non-invasive measurement of blood flow in human fetal great vessels and allows for retrospective motion correction and cardiac gating by making use of iterative reconstruction techniques. Fetal blood flow quantification has been performed with MRI in the past1,3,8,9. These studies had a prospective approach to mitigate motion corruption whereby scans would be repeated if gross fetal motion was visually identified from an initial reconstruction on the scanner. The current protocol improves on this by retrospectively rejecting data corrupted by gross fetal motion and further corrects for in-plane displacements arising from subtle fetal movements or maternal respiratory motion.
This protocol makes use of a multiparameter model for MOG whereby the RR interval for each fetal heartbeat is calculated. Using a low parameter heart rate model (such as 2 parameters) is generally acceptable for short scans since the healthy fetal heart rate has a low variability33. However, low parameter models become problematic for longer scans or in cases of pathologies such as arrythmia. A multiparameter model in MOG can track these changing RR intervals, providing more accurate flows.
The current protocol allows for some modifications. First, third party software used in this study for motion tracking and flow analysis can be replaced by other available software packages. Second, the number of iterations in the conjugate gradient descent algorithms for CS can be increased. In this study, the number of iterations in each step was set at a value beyond which there were minimal improvements based on prior reconstructions. In this work, only third trimester pregnancies were scanned. In earlier pregnancies, the fetus is smaller and there may be more room for motion. However, since quiescent periods in the scan are identified retrospectively for CINE reconstructions, this protocol should be successful for flow imaging at these earlier ages. An increase in the resolution of the scans may be required to cater to smaller vessel diameters at lower gestational age. For this protocol, the reconstruction times reported in Figure 1 and the Results are heavily dependent on the computational power available. For example, with better GPUs and more powerful processors, reconstruction times can be significantly reduced.
The protocol has certain limitations. First, the quality of the CINE reconstruction depends on the amount of data rejected in the motion correction step. With increasing episodes of gross fetal movements during a scan, more data is rejected. Consequently, the resulting signal-to-noise ratio (SNR) in the CINE reconstructions will decrease. Low SNR increases the uncertainty in the velocity images34 and the resulting flow quantification. Performance will therefore improve with greater fetal quiescence. Second, the method depends on the definition of ROIs for motion correction and MOG. In the current implementation, this step is performed manually. We have found that the reconstruction is stable to small differences in ROI position but this process results in wait times between the data acquisition and CINE reconstructions (since there are two ROI placement steps between the three iterative reconstruction steps). This becomes more cumbersome when there is a large number of slices acquired. In future implementations of the protocol, ROI placement will be automated.
Currently, we are using the presented protocol in research studies with approval from the local ethics board. The protocol can also be used in cases in which motion is a potential problem during an MRI exam, such as in neonates or uncooperative subjects. Future directions of the method involve investigating spiral trajectories35,36, which provide more efficient sampling and a possibility for exploring real-time fetal flow.
The authors have nothing to disclose.
None.
elastix | Image Sciences Institute, University Medical Center Utrecht | Image registration software | |
Geforce GTX 960 | Nvidia | 04G-P4-3967-KR | |
gpuNUFFT | CAI²R | Non-uniform fast Fourier transform | |
MAGNETOM Prisma | Siemens | 10849583 | |
MATLAB | MathWorks | ||
Radial Phase Contrast MRI sequence | Trajectory modification of manufacturer's Cartesian Phase Contrast sequence | ||
Segment | Medvisio | Data analysis | |
VENGEANCE | Corsair | LPX DDR4-2666 |