Accurate aortic annulus (AoA) sizing is crucial for transcatheter aortic valve implantation planning. Three-dimensional (3D) transesophageal echocardiography (TEE) is a viable alternative to the standard multidetector row computed tomography (MDCT) for such assessment, with few automatic software solutions available. The aim of this study was to present and evaluate a novel software tool for automatic AoA sizing by 3D TEE.
One hundred one patients who underwent both preoperative MDCT and 3D TEE were retrospectively analyzed using the software. The automatic software measurements’ accuracy was compared against values obtained using standard manual MDCT, as well as against those obtained using manual 3D TEE, and intraobserver, interobserver, and test-retest reproducibility was assessed. Because the software can be used as a fully automatic or as an interactive tool, both options were addressed and contrasted. The impact of these measures on the recommended prosthesis size was then evaluated to assess if the software’s automated sizes were concordant with those obtained using an MDCT- or a TEE-based manual sizing strategy.
The software showed very good agreement with manual values obtained using MDCT and 3D TEE, with the interactive approach having slightly narrower limits of agreement. The latter also had excellent intra- and interobserver variability. Both fully automatic and interactive analyses showed excellent test-retest reproducibility, with the first having a faster analysis time. Finally, either approach led to good sizing agreement against the true implanted sizes (>77%) and against MDCT-based sizes (>88%).
Given the automated, reproducible, and fast nature of its analyses, the novel software tool presented here may potentially facilitate and thus increase the use of 3D TEE for preoperative transcatheter aortic valve implantation sizing.
Accurate AoA sizing is crucial for preoperative TAVI planning.
A novel software tool for automatic AoA sizing by 3D TEE is presented.
The software measurements were validated against standard manual MDCT values.
It showed very good to excellent agreement with MDCT, being highly reproducible.
This software may potentially facilitate the use of 3D TEE for preoperative TAVI sizing.
Acquired aortic stenosis (AS) is the most common valvular heart disease in the elderly. In recent years, transcatheter aortic valve implantation (TAVI) has become an effective and widely used treatment option for patients with severe AS and high operative risk, with recent studies finding it viable for intermediate-risk patients as well. Because no direct access exists during the intervention, sizing of the transcatheter heart valve (THV) must rely on preoperative measurement of anatomic variables of the diseased valve. Accurate preoperative measurement of the aortic annulus (AoA) using imaging is therefore key to prevent procedural complications.
Although two-dimensional imaging modalities, such as two-dimensional transthoracic echocardiography and transesophageal echocardiography (TEE), were originally considered the standard for such assessment, it is now largely accepted that three-dimensional (3D) imaging techniques, namely, 3D TEE and multidetector row computed tomography (MDCT), should be used instead. Such 3D-based measures, such as AoA area and perimeter, were shown to present superior accuracy and reproducibility, even more when considering the elliptical shape of the annulus. Although MDCT is at most centers the modality of choice for this cross-sectional AoA measurement, its use is not without limitations, with concerns being raised in patients at risk for contrast-induced nephropathy or in the presence of imaging artifacts. As a result, 3D TEE is usually used as a radiation- and contrast-free alternative, with some centers specializing in echocardiography and even using it as the primary preoperative modality.
Because 3D-based AoA measurements rely on the correct identification of the cross-sectional plane that intersects the hinge point of each aortic valve (AV) cusp using a multiplanar reconstruction, manual analysis of 3D data sets is time consuming and complex, prone to both intra- and interobserver variability. Thus, the development of automatic tools to extract these measurements has the potential to shorten analysis time while reducing the variability among experts. Although several commercial automated solutions exist for MDCT-based analysis, only a few are available for 3D TEE.
Recently, our group proposed a novel, fully automatic (FA) framework for AV sizing based on 3D transesophageal echocardiographic data sets for TAVI planning, which was later embedded in custom noncommercial software, Speqle3D (KU Leuven, Leuven, Belgium). The latter allows integrated analysis and visualization of the results, with the user able to interact throughout the measurements’ extraction process. Although the framework was originally demonstrated to be accurate and robust against manual quantification on 3D transesophageal echocardiographic images, significantly reducing variability and analysis time, its accuracy was never assessed against MDCT-based measuring and THV sizing. Thus, in this study, we sought to perform such a comparison, while validating the feasibility and reproducibility of the measurements taken with this novel software tool. The software’s performance was further compared against a manual assessment from 3D transesophageal images. Finally, because the user can choose to perform a FA analysis or interact with it, both approaches were assessed and contrasted.
In the present study, high-risk or inoperable patients with symptomatic, severe AS who underwent TAVI at the St. Vinzenz-Hospital (Cologne, Germany) from August 2014 to September 2017 were retrospectively reviewed to identify those who underwent both preprocedural MDCT and 3D TEE. Among the 115 patients evaluated, 107 fulfilled these criteria and were thus further assessed. Of these, six patients were excluded because of inadequate preprocedural MDCT. No patients were excluded from imaging analysis on the basis of 3D transesophageal echocardiographic image quality or spatial-temporal resolution. As a result, 101 patients were enrolled in this study. All patients gave written informed consent before undergoing TAVI. Table 1 presents the clinical and echocardiographic characteristics of the study population.
|Age (y)||83.0 ± 5.1|
|NYHA class ≥ III||96 (95.1%)|
|Surgical risk assessment|
|Logistic EuroSCORE, % ( n = 18)||17.4 ± 15.0|
|EuroSCORE II, % ( n = 70)||5.5 ± 5.1|
|Porcelain aorta||2 (2.0%)|
|Creatinine > 2 mg/dL||6 (5.9%)|
|Cardiac risk factors|
|Cardiac surgery||7 (6.9%)|
|Percutaneous coronary intervention||45 (44.6%)|
|Preexisting pacemaker/defibrillator||14 (13.9%)|
|Myocardial infarction||13 (12.9%)|
|Cerebrovascular disease||5 (5.0%)|
|Peripheral vascular disease||8 (7.9%)|
|Chronic obstructive pulmonary disease||9 (8.9%)|
|Atrial fibrillation||37 (36.6%)|
|Mean transaortic gradient, mm Hg ( n = 96)||47.4 ± 17.4|
|AV area, cm 2 ( n = 98)||0.72 ± 0.19|
|mPAP, mm Hg ( n = 70)||32.8 ± 10.3|
|EF, % ( n = 100)||54.5 ± 14.6|
|Calcification severity moderate or greater ( n = 99)||77 (68.1%)|
THV Sizing and AV Implantation
On the basis of a thorough preoperative planning performed by a team of expert cardiologists and cardiothoracic surgeons, patients underwent transfemoral implantation of a balloon-expandable Edwards SAPIEN 3 THV (99 patients; Edwards Lifesciences, Irvine, CA) or a self-expandable Medtronic CoreValve Evolut R THV (two patients; Medtronic, Minneapolis, MN), of one of the three available sizes. At that time point, the experts’ team used AoA measurements from the available imaging modalities for THV sizing. In borderline cases (i.e., measurements between the manufacturer’s recommendations for two THV sizes), other clinical factors, such as gender, body size, and amount and location of calcifications were also considered. Overall, 23-, 26-, and 29-mm THVs were implanted in 44, 36, and 21 patients, respectively. All implantations were successfully performed, with the exception of one patient in whom the intervention was canceled because of increased risk for vessel rupture as a result of a too narrow iliofemoral artery.
Multidetector Row Computed Tomography
MDCT was performed according to guideline recommendations, using a multidetector 64-channel scanner (Discovery CT750 HD; GE Healthcare, Waukesha, WI). In short, the protocol included a prospective electrocardiographically triggered scan of the aortic root and heart acquired during a single inspiratory breath-hold. Tube potential and current were determined on the basis of patients’ body habitus. A timing bolus protocol was used to determine optimal contrast transit time. All images were reconstructed in an early diastolic phase using a soft tissue convolution kernel.
All patients underwent preprocedural TEE (Vivid E9 or E95 with 4V-D transducer; GE Healthcare). Besides two-dimensional standard imaging, the echocardiographic study included at least one 3D image of the AV and root acquired from a midesophageal position using the 3D zoom mode (i.e., displaying a smaller magnified pyramidal volume). In all images, the zoomed box covered the entire AV apparatus, from the left ventricular outflow tract (LVOT) to the sinotubular junction (STJ). At least one complete cardiac cycle was captured (either with or without stitching), with image resolution and size ranging from 0.31 to 0.69 mm and 162 × 162 × 162 to 325 × 325 × 201 voxels, respectively. Data sets were digitally stored in a raw-data format and exported to a workstation equipped with EchoPAC software (GE Healthcare). To enable offline automatic processing with our proposed software tool, the images were then anonymized, exported into an externally readable format, and converted to isotropic voxel spacing.
In cases with more than one 3D data set available, the image quality of each image was qualitatively assessed (on the basis of the overall image resolution and contrast and the presence and location of shadowing and stitching artifacts, among other visual aspects), and the best one was selected for the subsequent processing.
Manual Image Analysis
Multidetector Row Computed Tomography
The analysis of each cardiac multidetector row computed tomographic scan was performed by an experienced cardiologist at the time of TAVI planning, blinded to 3D transesophageal echocardiographic image analysis, using 3mensio Structural Heart software (3mensio Medical Imaging BV, Bilthoven, The Netherlands). In short, for each enrolled patient, after automated reconstruction and segmentation of the aortic root, the expert identified the cross-sectional image plane passing through the AoA plane (i.e., a virtual plane that simultaneously passes through the hinge point of the three cusps), which was then manually delineated to extract annular area.
During preoperative planning, 99 of the 101 enrolled patients had AoA manually measured in 3D transesophageal echocardiographic data also. In short, for each case, an experienced cardiologist evaluated the 3D data sets in commercially available software (EchoPAC), blinded to both multidetector row computed tomographic and automated 3D transesophageal analyses. At a midsystolic frame, the expert freely explored the volume using three orthogonal planes and positioned the short-axis view to intersect the aortic root at the level of the annulus. Once the plane was defined, AoA area was measured through manual delineation.
Automatic 3D Transesophageal Echocardiographic Image Analysis
Automatic AV analyses were performed using custom, noncommercial, MATLAB-based (MathWorks, Natick, MA) software, Speqle3D, in which our FA AV segmentation framework has been integrated ( Figure 1 ). In brief, the framework delineates the AV tract and extracts the AV measurements automatically, in a three-step approach: (1) AV detection in a transesophageal echocardiographic image volume at a given frame, (2) 3D AV segmentation, and (3) measurement extraction through the identification of the relevant short-axis planes ( Figure 2 ).
The first step focuses on delimiting the AV tract region to be subsequently segmented. Hereto, the image edge content is combined to the known vessel-like tube shape of the AV to detect the valve’s main axis, plus one point on either side of the valve (one at the LVOT and another at the ascending aorta). Starting from this initialization, the second step applies a shape-based segmentation strategy to three-dimensionally delineate the AV tract wall. Because shadowing artifacts are typically found in 3D transesophageal echocardiographic images from patients with AS, both intensity- and shape-based features are used in the automatic border detection algorithm. Shape-prior information is included using a statistical shape model (i.e., a model that statistically describes AV wall shape variability among the population), which guarantees that a valve-like shape is extracted and thus complementing the image content when in potentially misleading regions. The model was created from manually delineated valves, using transesophageal data from 20 patients not included in this study. The third step focuses on estimating the location of the relevant planes to be measured, namely, the LVOT, AoA, sinuses of Valsalva, and STJ. In this sense, known reference AV wall surfaces (in which these planes have been manually identified) are aligned to the AV wall surface of the target patient, allowing robust estimation of the planes’ location and orientation for the new image. By slicing the surface at each plane, the AV measurements (area, perimeter, or diameters) are extracted. See the original publications for details.
Within Speqle3D, after loading the patient’s volume sequence and identifying the midsystolic frame of interest, the user can sequentially apply each of the aforementioned steps to ultimately extract the annular area. In addition, the Speqle3D environment allows the user to interact, if needed, with the result of each step ( Figure 3 ), being thus able to (1) correct the valve’s axis or the location of the two points placed along it; (2) refine the delineated AV wall surface by manually adding control points in suboptimal regions (the surface is then nearly instantaneously corrected by automatically fusing this user input with the automatic border detection algorithm, similar to Barbosa et al. ; note that such fusion guarantees that a valve-like shape is still extracted, instead of simply creating a deformed version of the surface that passes through the user points); and (3) rectify the location and/or orientation of the identified short-axis planes.
In the present study, because of their clinical value, only annular measurements were assessed.
FA versus Interactive Analysis
To compare the results obtained with the proposed automatic framework with or without user interaction, an observer performed the analysis twice for each patient, in a random order and blinded to the results of each analysis. In the first analysis, the observer simply loaded the image, chose the midsystolic frame, applied all automatic steps consecutively, and then registered the measurements (FA analysis). In the second analysis, and assuming the same mid-systolic frame, the observer was allowed to interact with the Speqle3D’s user interface in order to correct the results, if deemed necessary, after each step (henceforth referred to as interactive analysis [IA]). For the sake of completeness, two additional analysis were also performed with intermediate levels of user interaction (i.e., correction allowed only after first or first and second steps), whose description and results are included in Appendix 1 (available at www.onlinejase.com ).
Intra- and Interobserver Analysis
To establish the intraobserver reproducibility of the proposed software, the same observer repeated the aforementioned analysis after 15 days. Moreover, a second observer also performed the analysis to assess interobserver variability. Note that each was blinded to the other’s results. All analyses considered the same midsystolic frame.
To further evaluate the robustness and reproducibility of the proposed software, a test-retest analysis was also performed. Enrolled patients with more than one 3D transesophageal echocardiographic image sequence were gathered (in a total of 65 cases), and a second 3D data set was selected from each (on the basis of the overall image quality assessment). The selected 3D sequences were then analyzed by the first observer (with a 3-week interval from the first analysis). Both FA analysis and IA were performed, with the user able to choose the midsystolic frame for these data sets at will.
Theoretical Prosthesis Size Selection
To assess the potential effect on TAVI of the measurements taken (with manual analyses or by the novel automatic 3D transesophageal echocardiographic software), the theoretical prosthesis size was identified for each patient and each measuring approach. To this end, manufacturer-recommended sizing algorithms were considered ; this experiment was applied to all patients implanted with Edwards SAPIEN 3 THVs (99 of 101). Because these recommendations were originally created on the basis of multidetector row computed tomographic area measures, any statistically significant systematic difference (i.e., the bias) observed between a 3D transesophageal echocardiographic measuring approach and MDCT was corrected for. In other words, 3D transesophageal measures were corrected on the basis of the observed bias with respect to multidetector row computed tomographic values, and the MDCT-based sizing rules were then applied. Moreover, because the recommendations present an overlap between the three THV sizes, allowing over- or undersizing according to patient-specific characteristics, borderline cases (i.e., whose measures lay in the overlapping regions) were also considered, leading in the end to five groups (23, 23/26, 26, 26/29, and 29 mm).
Evaluation Approach and Statistical Analysis
The feasibility of the FA measurements extraction was assessed by quantifying the number of successfully analyzed cases with respect to the total number of patients enrolled. Because applicability depends mostly on the correct detection of the AV tract, an analysis was considered successful, upon visual inspection, if the detected and subsequently segmented tract roughly spanned from the LVOT to the ascending aorta. Only successfully analyzed data sets were then considered in the FA results. In opposition, an IA is always feasible, as the user can refine the AV axis before the subsequent processing.
The accuracy of the automatically computed 3D transesophageal echocardiographic area measures was assessed by comparing against manually determined values on MDCT and 3D TEE. The measured area values are given as mean ± SD for each measuring approach. To assess agreement, Bland-Altman (BA) analyses were performed between automatic and manual measures, computing both biases and limits of agreement (LOA; i.e., 1.96 × SD). The agreement between manual approaches was also calculated. The statistical significance of the biases was tested using a two-tailed t test against zero (significance at P < .05), while a two-tailed F test was used to compare the LOA. Moreover, the intraclass correlation coefficient (ICC) and its 95% CI were also computed using SPSS Statistics version 20 (IBM, Armonk, NY). Consistency agreement was considered for measures taken from different modalities, while absolute agreement was used when extracted from the same modality.
To quantify the differences obtained when interacting or not with the software (FA and IA, respectively), the same agreement analysis was performed between these measuring approaches. Moreover, intra- and interobserver variability, as well as test-retest variability, were evaluated to assess the software’s reproducibility, calculating agreement between repeated measures. For the sake of completeness, the ICC (and associated 95% CI, considering absolute agreement) and the coefficient of variation (CV; i.e., the SD of the differences divided by the average of all measured values) were also computed.
The total time taken to perform an analysis (either FA or IA, starting after loading the volume and until finishing the measures’ extraction) was recorded.
Finally, to assess the clinical significance of the extracted measures for TAVI planning, the sizing agreement between the theoretically selected THV size and the implanted one was evaluated. Borderline cases were considered successfully sized if either of the possible THV sizes was the implanted one (e.g., successful for the 23/26 range if either a 23- or a 26-mm THV was implanted). Besides comparing each measuring approach against the implanted size, the sizing agreement between pairs of approaches was also assessed. Such comparison enables understanding of the actual effect of the differences between measures, disregarding the TAVI team’s specific decisions during the original planning and/or intervention. Once again, any borderline case was considered successfully sized if it matched the other approach’s size (or sizes, if also in the overlapping region).
On the basis of the aforementioned visual criteria, the FA analysis was found to be feasible in 92.1% of the cases (93 of 101). The unsuccessfully analyzed cases were related to a suboptimal detection of the valve’s main axis, which led to an incorrect delineation of the AV wall. For these cases, during IA, the user corrected the axis positioning, with the following steps leading to an accurate assessment. IA was feasible in all cases.
Three-Dimensional TEE versuss MDCT
Table 2 presents area measures for both automatic 3D transesophageal echocardiographic approaches (FA and IA) and manual ones, and summarizes the agreement between them. Both automatic approaches showed very good agreement against manual measures on MDCT and 3D TEE, with the IA having slightly narrower LOA. Nonetheless, there were no statistically significant differences between the LOA of both approaches. Between themselves, the FA analysis and IA had excellent agreement (BA, 8.4 ± 35.2 mm 2 ; ICC = 0.89 [95% CI, 0.83–0.92]). Importantly, when splitting the study population on the basis of the calcification severity (none/mild vs moderate/severe cases; Table 1 ), no significant differences were observed between the LOA obtained against manual measurements (see Appendix 2 ; available at www.onlinejase.com ). Figure 4 illustrates representative segmentation results for three patients.
|Method||n||Area (mm 2 )||Comparison with MDCT MA||Comparison with 3D TEE MA|
|BA||ICC (95% CI)||BA||ICC (95% CI)|
|MDCT MA||101||468.2 ± 89.5||—||—||58.5 ∗ ± 51.8||0.81 (0.73–0.87)|
|3D TEE MA||99||409.7 ± 79.4||−58.5 ∗ ± 51.8||0.81 (0.73–0.87)||—||—|
|3D TEE FA||93||423.5 ± 76.0||−45.4 ∗ ± 54.7||0.78 (0.69–0.85)||15.6 ∗ ± 54.8||0.74 (0.62–0.82)|
|3D TEE IA||101||415.4 ± 76.2||−52.8 ∗ ± 48.0||0.83 (0.76–0.88)||6.8 ± 46.0||0.82 (0.75–0.88)|
Intra- and Interobserver Variability
Because the FA analysis does not depend on the user, no intra- or interobserver variability exist. In its turn, IA was found to be highly reproducible, in terms of both intraobserver (BA, 0.3 ± 20.0 mm 2 [i.e., bias ± SD]; ICC = 0.97 [95% CI, 0.95–0.98]; CV = 4.8%) and interobserver variability (BA, 14.0 ± 23.0 mm 2 ; ICC = 0.94 [95% CI, 0.85–0.97]; CV = 5.6%). Interestingly, similar reproducibility was found irrespective of the level of user interaction applied (see LOA and ICC values in Appendix 1 ; available at www.onlinejase.com ).
Table 3 presents the area measures extracted for the cases with two 3D transesophageal echocardiographic sequences and summarizes the agreement between studies for both automatic approaches. Once again, the software showed highly reproducible measurements (ICC > 0.93 for both types of analysis).
|Method||n||Area S1 (mm 2 )||Area S2 (mm 2 )||Test-retest variability|
|BA||ICC (95% CI)||CV|
|3D TEE FA||59||420.4 ± 67.0||423.4 ± 78.3||−3.4 ± 27.8 †||0.93 (0.89–0.96)||6.6%|
|3D TEE IA||65||403.5 ± 67.5||390.1 ± 71.7||13.4 ∗ ± 20.5||0.94 (0.83–0.97)||5.2%|
The FA approach took, on average, 19.0 ± 1.9 sec, while the IA approach required 63.0 ± 11.6 sec. Note that this time already includes frame selection, plus the computation of all measurements. In Appendix 1 (available at www.onlinejase.com ), the analysis time for intermediate levels of user interaction is also provided. In contrast, notice that a routine measuring analysis on 3D TEE takes about 5 min.
Valve Sizing Agreement
The distribution of THV sizes among the study population for each approach, plus the clinically implanted valve, is illustrated in Figure 5 . The sizing agreement for each against the implanted prosthesis is given in Table 4 . The agreement against manual measures on MDCT and 3D TEE is also presented. The highest agreement against implanted THV sizes was found for the manual multidetector row computed tomographic analysis, which is primarily linked to the fact that the original planning was based primarily on MDCT-based measurements. In its turn, all 3D transesophageal approaches (manual, FA, and IA) showed very similar sizing agreements, presenting better agreement when directly comparing against the theoretically selected multidetector row computed tomographic sizes. Between themselves, the 3D transesophageal automated analyses (FA and IA) agreed in 97.8% of the cases.