We study multimodal images acquired from 5 patients with hypoplastic left heart syndrome (HLHS), 3 post-Norwood 1, and 2 post-Hemifontan, with an age range of 0.18–3.40 years, and weight in the range of 4.93–15.4 kg. These patients underwent a research ultrasound examination immediately after the clinical MR examination, both under general anesthesia (GA). Transthoracic ultrasound volumes were acquired using a Philips iE33 system and a cardiac X5-1 3D transducer, from subcostal windows.

Cardiac magnetic resonance (CMR) imaging was performed using a 1.5 T MRI scanner (Philips Intera Achieva, Philips Healthcare, Best, Netherlands). RV volumes and function were obtained as part of a comprehensive functional evaluation. In accordance with our unit protocol for CMR evaluation of HLHS, a single stack of contiguous 6–8 mm balanced SSFP slices (TR 1.8 ms, TE 3.5 ms, FOV 180–320, 40 phases per cardiac cycle, 6–12 lines per segment depending on heart rate, acquired resolution 1.2 mm × 1.2 mm to 1.8 mm × 1.8 mm) oriented in a plane equivalent to the short axis of the tricuspid valve were obtained in an end-expiratory breath-hold of 4–7 s per slice.

This study was carried out in accordance with the principles of the Declaration of Helsinki. Ethical approval was granted by the local ethics committee “Advanced Echocardiography in Pediatric Patients” at Guy’s and St. Thomas’s and King’s College London (09/H0802/116) after informed consent was obtained from the patients parents.

Segmentations on both MR and ultrasound images were done manually. Semiautomatic methods based on processing of the image data, such as model based segmentation, level sets, region growing, and other methods can introduce a bias in the comparison because MR and ultrasound images perform very differently on them. Consequently, we have used the manual segmentation tool provided by the MITK software (6) to segment both modalities. The segmentation was carried out by contouring the endocardium on a stack of short-axis planes and then interpolating the contours to form a volume.

Echo to CMR alignment (registration) is a challenging problem. Image features are normally not consistent between the two modalities because not all structures that are visible in the ultrasound image (e.g., trabeculae, valves) are also visible in the CMR image. Moreover, view-dependent artifacts characteristic to ultrasound imaging (e.g., shadows, reverberations) lead to erroneous image features that obviously are not present in the CMR image. Last, structures that are visible in both modalities are captured in a very different way: for example, ultrasound image formation can cause thickening or narrowing of these structures depending on the angle of incidence of the ultrasound wave, while no such view dependency takes place in CMR image acquisition. As a result, most automated and semi-automated, image- or feature-based registration algorithms fail to align CMR and echo images accurately.

For this study, we have carried out image alignment by manually selecting corresponding ventricular landmarks in both modalities and calculating the rigid transform (rotation and translation) between the two landmark sets. An independent operator carried out registrations for all patients to ensure that the same alignment was used when comparing segmentations carried out by different experts.

Landmark selection is done as follows: first, the base-to-apex axis was found (Figures 1A,B). Along this axis, a point at mid-height of the ventricle (represented in the figure by a white dot) is selected to produce a short-axis slice (Figure 1C). On this short view, the in-plane axes are rotated and translated so that one plane is parallel to the diaphragm, and the other passes by the closest papillary muscle (Figure 1D) while maintaining the slicing planes orthogonal to each other.

Without changing the orientations of the slicing planes, the crosshair is translated following the through short-axis direction to the atrioventricular valve plane (Figures 1E,F). At that location, six landmarks are selected: the center of the atrioventricular valve (1), the inferior (2), anterior (3), left (4) and right (5) sides of the valve annulus, and the ventricular apex (6). Following the same process, corresponding landmarks are selected in the CMR image (Figures 1G,H). The rigid transformation between the two point sets was found using the least squares method described by Arun et al. (7).

Figure 2 shows the image registration results for 5 patients. A superimposition of the medial short-axis slice from both CMR and echo images, for each patient, is shown. A selection of movies showing the achieved alignment and its consistency over time are included in Supplementary Material.

In order to carry out a regional analysis of the difference in estimated volume, we need to divide the ventricular volume into sectors. There is little literature on regional analysis of the right ventricle (RV) in Fontan patients like those with hypoplastic left heart syndrome (HLHS). For repaired Tetralogy of Fallot (ToF) patients, Zhong et al. (8) proposed a 15-segment subdivision.

This subdivision is, however, not well suited for single ventricle circulation because of the essential differences in RV morphology between the two cases. In repaired ToF, the RV morphology is not very different from a normal RV. In Fontan circulation, there is no functional left ventricle (LV), hence the RV supports the systemic circulation and has adapted its morphology becoming more globular, toward the shape of a normal LV.

For this reason, other means of describing the RV anatomy in a standardized way have been proposed. Menon et al. (9) carried out regional analysis from 2D echocardiography by dividing the myocardium in a parasternal long-axis view into four sections and in a four chamber view into 6 segments. Wong et al. (10) carried out a 3D analysis on RV morphology and function in HLHS using a population-based atlas, which defined ventricular anatomy with respect to the position of the LV remnant. This representation allowed the estimation of regional strain dividing the RV into basal, medial, and apical layers.

Inspired by the divisions carried out by Menon et al., Zhong et al., and Wong et al. (8–10), and taking into account the standard AHA 17 segment division for the LV proposed in Ref. (11), in this paper, we propose to divide the 3D RV shape into 14 segments as indicated in Figure 3. Along the RV axis, four layers are defined: apical, medial, basal, and valvular. The apical layer consists of a single sector. The medial layer and the basal layer are divided into 4 and 6 sectors, respectively (similar to the LV AHA division). The valvular layer is divided into 3 sectors that cover half of the circle reflecting the asymmetric shape of the RV, and cover the “shoulder” area underneath the inflow valve. This layer covers the upper anterior wall (1), the upper lateral free wall (2), and upper inferior free wall (3). The basal layer, situated below the valvular layer, starts at the intersection point with the anterior (4), lateral (5), and inferior (6) parts of the free wall and the inferoseptal (7), lateral-septal (8), and anteroseptal (9) sectors. The medial layer includes four sectors covering the anterior (10), the lateral free wall (11), the inferior (12), and the septal (13) sectors. The apical layer includes a single sector (14). Figure 3A shows a representation of a RV for reference. Figure 3 shows a 3D representation of the segment division for one patient from the echo-derived segmentation and from the CMR-derived segmentation. Colors are matched by the bulls-eye plot diagram in Figure 3D, which is used as model for the results in the remainder of this paper. The orientation of the bulls-eye plot and the denominations “anterior” (A) and “inferior” (I) are consistent with that in Ref. (11). We have replaced the septal and lateral names in (11) by right (R) and left (L), because we believe this is a more straightforward and intuitive notation in systemic RV patients.

An advantage of the landmark selection process described in Sec 2.3 is that it allows to define the segment division automatically since the RV axis is defined by points 1 (center of tricuspid valve) and 6 (RV apex), and the superior-to-inferior direction is defined by the points 2 and 3.

Three experts carried out EDV segmentations on 5 pairs of pre-aligned MR and ultrasound images, as described in Sec. 2.2. These images were spatially aligned as indicated in Sec. 2.3. The resulting aligned segmentations were divided into 14 segments as described in Sec. 2.4.

The resulting end diastolic volumes (EDVs) are compared globally (ΔEDV = EDV_echo − EDV_CMR) in absolute terms, and also regionally. For the regional analysis, the difference between echo-derived and CMR derived regional volumes is expressed as a fraction of the total EDV volume estimated from CMR: (1)ΔEDVr=EDVr,echo−EDVr,CMREDVCMR,for every region r

This allows us to compare the obtained values across patients. In order to compute statistics on segmentations from multiple experts, the average segmentation was computed by averaging the binary masks representing the EDV segmentations followed by a thresholding operation where voxels with an intensity greater than 0.5 were kept.

In addition to the numerical analysis, we have carried out a qualitative analysis by comparing the average segmentation contours at four different short-axis planes uniformly spaced along the RV axis for each patient.

Global EDV differences are shown in Table 1. The numbers reflect the average ± SD over all experts, for each patient and each modality, in milliliters. The numbers reported are within the normal range and variability to other studies in the literature as reported by Simpson et al. (5).

Large variability in the CMR derived volumes in patient 1 are associated to low echo image quality (as shown in Figure 6, top left), which led to large differences on how experts decided to include some structures like the papillary muscles.

The results of the regional analysis of EDV differences are shown in Figures 4 and 5. Figure 4 shows the average relative difference in regional EDV bulls-eye plot for each patient, in percentage. There is a common pattern across all datasets where the highest disagreement is near the apex, decreasing gradually near the valve plane.

Figure 5 shows the integrated results from all patients as a bar chart. This representation has been chosen instead of the bulls-eye plot in order to accurately show the average values as well as the SD as error bars. The sector number corresponding to each bar is indicated next to the bar. Note that, in the basal layer, the order of the bars has been modified so that sectors that are approximately aligned vertically are represented as bars that are aligned horizontally.

The results shown in Figure 5 are consistent with the results shown in Figure 4E. The highest average disagreement takes place at the apex, with a −14.2% relative difference between echo-derived volume and CMR-derived volume. The spatial distribution of the error in the medial layer is uniformly distributed and close to −5% in average. In the basal and the valvular layers, the difference between the two modalities is significantly smaller. Interestingly, in sectors near the outflow tracts (1, 2, 4, and 5), there is a high variance across patients and operators. The qualitative results in the next section expand on the potential reasons for this.

Figure 6 shows a selection of short-axis, end-diastole views of both CMR and echo including the outline of both echo and CMR-derived segmentations (after alignment), to illustrate the most significant findings of this study. The ventricular segmentation from echo is represented as a red contour and the CMR-derived segmentation as a green contour. A comprehensive collection of short-axis views for all patients at different planes is included in Supplementary Material.

The 2D slices in Figure 6 illustrate sources of systematic differences between EDV segmentations carried out using CMR images and echo images. On the first column on the left, a very significant difference between the two contours can be observed. Echo-derived segmentation was particularly challenging in this patient due to the low-image quality (compared to images from other patients). Additionally, the echo-derived contour was drawn excluding trabeculae from the segmentation. In the CMR image, where the visibility of the endocardium is poor, the segmentation appears to run closer to the myocardium.

On the second column from the left, it can be seen that image quality is relatively high in both modalities, but the difference comes from the contours delineating different structures. The echo-derived contour (in red) correctly follows an image edge (top row), and the CMR-derived contour (in blue) also follows an edge in the CMR image (bottom row). The edge in the echo image represents the trabeculations, which are not visible in the CMR image, where the contour follows the myocardium. A similar effect can be observed in the third column (patient 3).

The fourth column shows the effect of lack of boundary definition in the ultrasound image on the resulting segmentation. In this case, overall image quality is high, but the anterior wall is not visible due to shadowing from the air in the lung. As a result, the delineated contour does not match the real structure, which is visible in the CMR image. This finding is reflected in our proposed bulls-eye plot in Figure 4D.

The fifth column illustrates disagreement due to lack of boundary definition. An unclear boundary was delineated as a true boundary on the echo image, while it was considered part of the ventricular cavity in the CMR.