Between June 2014 and January 2015 adult patients planned for clinical contrast-enhanced CMR were consecutively approached to undergo the supplemental CMR 4D flow examination, and prospectively included in the study (Table 1). The study was carried out in accordance with the principles of the Declaration of Helsinki and approved by the institutional Medical Ethics Committee. All participants gave written informed consent.

The acquisition was performed using a 1.5 T or 3.0 T whole body scanners (Discovery MR450 and Discovery MR750, 45 mT/m, 200 T/ms, GE Medical Systems, Milwaukee, WI, USA) with a software version DV24.0, using dedicated 32-channel phased-array cardiac surface coils. Depending on the clinical request a variety of CMR examinations were performed including multi-planar cine steady-state free precession acquisitions, flow visualization using phase contrast sequence, contrast-enhanced angiography or delayed enhancement imaging.

The CMR 4D flow data was acquired immediately, or shortly after the bolus injection of 0.1–0.2 mmol/kg gadolinium-based contrast agent (Gadovist 1 mmol/ml, Bayer, Mijdrecht, The Netherlands), depending on the clinical indication for contrast administration. The sequence was prescribed in the axial plane, with a matrix of 192 × 144 and 2.8 mm slices interpolated to 1.4 mm slices, while the entire thorax was included in the field of view during a free-breathing, ECG gated scan. Patients were also asked to breath regularly during the acquisition. The median views per segment was 4 and the median repetition time was 3.9 ms, resulting in a temporal resolution of 63 ms. When necessary, the views per segment and temporal resolution were adapted to limit the scan duration to approximately ten minutes (Table 2). Following recommendations from literature, the velocity encoding value was set at 250 cm/s, which is a good compromise for the 4D flow to allow acquiring low and high velocities simultaneously [11].

Because CMR 4D flow imaging is a phase contrast sequence, similar to planar flow acquisitions, magnitude images for anatomic information as well as images containing flow information may be reconstructed from the data. These large unreconstructed raw data sets (approximating 5 Gb per patient) were uploaded to a dedicated web-based software application (Arterys Inc., San Francisco, CA, USA) for data correction, visualization of the anatomical and flow components, and evaluation of the aortic regurgitation. As previously described [6], images were reconstructed for each cardiac temporal phase with a combined autocalibrating parallel imaging compressed sensing algorithm (L1-SPIRIT). Data was processed using large servers (Amazon Inc., Seattle, WA, USA). They are equipped with high frequency processors (Intel Xeon E5-2670, Intel Corporation, Santa Clara, CA, USA) and high-performance graphics processing units (NVIDIA Corporation, Santa Clara, CA, USA). The user logged into the application via internet connection from any type of computer or other device. Before the data sets were analyzed, the user could semi-automatically correct the images for encoding errors related to gradient field distortions and eddy currents. This approach is based on defining a threshold to identify regions with static tissue. These were then used to estimate eddy current induced varying phase offset errors, which were subsequently subtracted from the entire image [12]. Once the post-processing was complete and eddy currents correction performed the software facilitated interactive, real-time, manual navigation through the imaging volume using perpendicular multiplanar reformations and volume rendered reconstructions. The data could be reformatted into planes representing specific regions of interest (Fig. 1; Video 1). Flow could be displayed by color-coding of the velocity information or by vector rendering to display flow direction. For a red color on the color scale, a treshold of 150 cm/s was set.

The aortic valve was localized within the three dimensional anatomical dataset and displayed on three cross-sectional views that are perpendicular to each other. Using the cine display mode with superposition of flow onto the anatomy, the data was screened for the presence of reversed flow from the aortic valve during diastole. Identification of suspected aortic regurgitation was then evaluated using a temporally frame-by-frame approach. Because the velocity-based color-coded display of flow lacks directional information, the vector arrow display was used for confirmation (Fig. 1). Aortic valve regurgitation was defined as mild, moderate or severe. The grading criteria were adapted from echocardiographic recommendations [13, 14] using only the parameters available for both echocardiography and CMR 4D flow imaging: the ratios between the width of regurgitant jet and of the left ventricle outflow tract, and between the length of the regurgitant jet and of the left ventricle and the presence of reversal flow during diastole at the level of the descending aorta (Table 3). Two blinded readers (RGC and KN with 4 and 7 years of CMR experience, respectively) independently assessed each case and graded aortic regurgitation. A joint consensus reading served to solve discordant interpretations.

Two-dimensional transthoracic Doppler echocardiography was performed using a commercially available system (IE33, Philips Medical Systems, Best, The Netherlands). The gain settings were adapted to the patient’s characteristic to obtain the best quality of color Doppler. Because usually in aortic regurgitation the velocities are higher than 2 m/s, it is not possible to correct for aliasing phenomenon even by changing the aliasing velocity settings. Thus practically aliasing is always present. We set the aliasing velocity at 1.2 m/s.

Blinded to the CMR 4D flow imaging results, echocardiography images were independently evaluated by two cardiologists (AvdB and JR-H with 15 and 23 years of echocardiography experience, respectively), followed by a consensus reading. The presence of aortic regurgitation was graded using the same criteria as described above for the CMR 4D flow imaging. Because our paper focusses on the technical validation of 4D flow imaging we only considered parameters that could be assessed by both techniques.

Statistical analysis was performed using SPSS (version 21 IBM, Armonk, NY, USA) and MedCalc (version 13.0; MedCalc Software, Ostend, Belgium). Categorical variables were reported as totals and percentages and continuous variables with a normal distribution as means ± standard deviations (SDs), or if data was skewed, as median with interquartile range. The diagnostic characteristics of CMR 4D flow imaging were evaluated against echocardiography by using C-statistics, including sensitivity, specificity, positive predictive value, negative predictive value with their corresponding 95 % confidence intervals (CIs), accuracy and Cohen’s kappa agreement test.

The correction and visualization of the CMR 4D flow data and interpretation of aortic regurgitation required 10–15 minutes per patient. The agreement between CMR 4D flow imaging and echocardiography for the grading of regurgitation was good (κ = 0.73; Table 4). Representative case examples are shown in Figs. 2, 3 and Video 2. By CMR 4D flow imaging all four patients with moderate and 9 out of 15 patients with mild regurgitation were correctly identified. One case of mild regurgitation by echocardiography was interpreted as moderate by CMR 4D flow, and in 1 out of 35 cases CMR 4D flow identified mild regurgitation while echocardiography showed none. For the detection of any regurgitation, CMR 4D flow imaging demonstrated specificity of 97 % (95 % CI: 85–99 %), sensitivity of 74 % (CI: 49–91 %), positive predictive value of 93 % (CI: 68–99 %) negative predictive value of 87 % (CI: 73–96 %), accuracy of 88 % and κ = 0.74. To identify clinically relevant, more than mild regurgitation, CMR 4D flow imaging had sensitivity of 100 % (CI: 40–100 %), specificity of 98 % (CI: 89–100 %), positive predictive value of 80 % (CI: 29–97 %) and negative predictive value of 100 % (CI: 92–100 %), accuracy of 98 % and excellent agreement with echocardiography κ = 0.88.

Echocardiography has been the primary imaging technique for the care of patients with structural heart disease, offering advantages in ease of performance, safety and cost. However, in complex disease and after surgery echocardiography alone is frequently insufficient. CMR offers an alternative imaging option that is helpful in a number of specific applications, e.g. imaging of myocardial perfusion or scar, flow quantification, or when echocardiography cannot be performed adequately due to technical reasons, such as insufficient acoustic window. By the current standard of care, CMR measures flow using two-dimensional, velocity-encoded sequences, and it displays flow perpendicular to a manually selected plane. It can be a lengthy process that requires good cooperation from the patient. Potential advantages of CMR 4D flow protocol are that anatomical, functional and flow information are obtained during a free-breathing acquisition of 7–10 minutes, which is more comfortable for the patient and easier to perform. Without a need to specify beforehand, or expert assistance during the examination, flow can be measured anywhere and in any direction within the great vessels of the thorax after the data has been acquired. While 2D sequences acquire flow information in a static plane, CMR 4D flow imaging allows for dynamic alignment of the plane of interest to the position of moving structures (e.g. valve annulus) [15]. Three-dimensional visualization of flow is also helpful for assessment of regurgitation jets that change direction during the heart cycle. However, clinical application of CMR 4D flow imaging is met by several technical challenges. Accurate representation of flow requires several pre-processing steps to correct for eddy currents and for magnetic field inhomogeneity. Experienced centers often rely on in-house developed solutions. To our knowledge this is the first report on the performance of a cloud-based software that applies volumetric eddy currents correction and direct, interactive flow visualization of CMR 4D flow data in a single tool package by internet connection, with remote processing of transferred raw datasets on high-performance workstations.

This is a relatively small, single-center study. As a proof of concept we limited the analysis to the qualitative evaluation of aortic valve regurgitation. Because demonstration of technical feasibility for flow visualization was the aim of this study, only parameters available for both echocardiography and CMR 4D flow imaging were used for assessment of the regurgitation severity. The prevalence of aortic regurgitation was 35 %, though none of the patients had severe regurgitation. Future studies should expand into quantitative analyses, as well as a wider range of valve pathologies.

Future work is needed to determine optimal technical parameters for quantification of stenosis, regurgitation, shear stress, pressure gradients and other flow field derived metrics. To improve the contrast to noise ratio, we performed the CMR 4D flow acquisition after clinically-indicated intravenous contrast agent administration. Future work may focus of determining diagnostic accuracy without the use of contrast agent.