Deep Learning-Based Fully Automated Aortic Valve Leaflets and Root Measurement From Computed Tomography Images　― A Feasibility Study ―

Haruo Yamauchi; Gakuto Aoyama; Hiroyuki Tsukihara; Kenji Ino; Naoki Tomii; Shu Takagi; Katsuhiko Fujimoto; Takuya Sakaguchi; Ichiro Sakuma; Minoru Ono

doi:10.1253/circj.CJ-24-1031

Abstract

Background: The aim of this study was to retrain our existing deep learning-based fully automated aortic valve leaflets/root measurement algorithm, using computed tomography (CT) data for root dilatation (RD), and assess its clinical feasibility.

Methods and Results: 67 ECG-gated cardiac CT scans were retrospectively collected from 40 patients with RD to retrain the algorithm. An additional 100 patients’ CT data with aortic stenosis (AS, n=50) and aortic regurgitation (AR) with/without RD (n=50) were collected to evaluate the algorithm. 45 AR patients had RD. The algorithm provided patient-specific 3-dimensional aortic valve/root visualization. The measurements of 100 cases automatically obtained by the algorithm were compared with an expert’s manual measurements. Overall, there was a moderate-to-high correlation, with differences of 6.1–13.4 mm² for the virtual basal ring area, 1.1–2.6 mm for sinus diameter, 0.1–0.6 mm for coronary artery height, 0.2–0.5 mm for geometric height, and 0.9 mm for effective height, except for the sinotubular junction of the AR cases (10.3 mm) with an indefinite borderline over the dilated sinuses, compared with 2.1 mm in AS cases. The measurement time (122 s) per case by the algorithm was significantly shorter than those of the experts (618–1,126 s).

Conclusions: This fully automated algorithm can assist in evaluating aortic valve/root anatomy for planning surgical and transcatheter treatments while saving time and minimizing workload.

The prevalence of degenerative aortic valve stenosis (AS) has increased due to longer life expectancy, and it has a poor prognosis if not treated appropriately.¹ Transcatheter aortic valve replacement (TAVR) is now a major treatment option for AS because it is less invasive and has substantial mid-to-long-term outcomes in low surgical-risk patients.²^–⁴ Aortic regurgitation (AR), often present with AS,⁵ can be caused by primary valve deformity and/or root dilatation (RD).⁶^,⁷ Recently, more patients have been opting for aortic valvuloplasty and valve-sparing root replacement (VSRR) as a viable alternative to prosthetic valves,⁸^,⁹ so knowing the geometric height (gH) and effective height (eH) of the aortic valve leaflets is crucial for the long-term durability of a spared valve by ensuring proper coaptation during valve closure.¹⁰^,¹¹

Precise knowledge of the aortic valve/root geometry is essential for planning catheter-based interventions and surgery, and although echocardiography can be used to assess the severity of valve diseases in real time, its performance is operator-dependent. ECG-gated cardiac computed tomography (CT), with its superior spatial resolution and objective image data, is particularly useful for estimating TAVR device selection,¹² and potentially for planning valvuloplasty or VSRR.⁶^,¹³ However, measuring the aortic valve/root geometry typically requires manual or semi-automatic methods, which are time-consuming and labor-intensive.¹²

To automate aortic root measurement from CT data, several studies have reported deep learning-based methods for TAVR planning;¹⁴^–¹⁹ however, detailed information on the aortic valve leaflets is lacking. To address this issue, we developed a deep learning-based fully automated aortic valve leaflet and root segmentation and measurement (AVRM) algorithm using CT images,²⁰ which provides patient-specific 3-dimensional (3D) aortic valve morphology with measurement data. However, that method could not accurately assess aortic RD and we have not validated the algorithm’s clinical applicability by comparing the segmentation results with current standard manual measurements.

Thus, this study aimed to (1) retrain our AVRM algorithm for RD; (2) reconstruct 3D views of the aortic valve/root using our updated algorithm; (3) assess the feasibility of our algorithm by comparing the automated measurement values with the manual measurement values obtained by 3 experts in this field; and (4) estimate its efficiency by comparing the assessment time between automated and manual measurements.

Methods

Data Collection

Cardiac 3D CT imaging data were retrospectively collected at the University of Tokyo Hospital, per the eligibility criteria outlined in the following section. Most CT images, except for those examined in other institutions, were acquired with contrast enhancement using Aquilion ONE (Canon Medical Systems Corporation, Otawara, Tochigi, Japan). The scanning parameters were: detector configuration, 320×0.5 mm; tube potential, 120 kV; tube current, 250–760 mA depending on body habitus. Patients received 22.2 mg/kg/s of Iopamidol 370 mg/mL (Iopamiron 370: Bayer, Osaka, Japan) with a mean administered volume of 71.0±17.1 mL (range, 30.5–97.0 mL) over 14 s. The ECG-gated cardiac CT sequences included 1–11 volumes per cardiac cycle, where each volume contained 181–1,011 slices with 512 pixels. The in-slice resolution was isotopic with 0.29–0.43 mm, and the slice thickness was 0.25–0.5 mm. The volume data sets at end-systole and end-diastole (30% and 80% of a cardiac cycle, respectively) were transferred to the image processing system Vitrea workstation²¹ (Canon Medical Informatics, Inc., Otawara, Tochigi, Japan) for the ensuing AS and AR/RD analyses, respectively.

This study was conducted per the guidelines of the Declaration of Helsinki, and was approved by the Research Ethics Committee of the Faculty of Medicine, The University of Tokyo, on January 27, 2017 (approval number: 11330). The requirement for written informed consent was waived by the Institutional Review Board because patient data were retrospectively collected, anonymized, and de-identified before use in this study.

Eligibility Criteria for the Analyses of AS and AR/RD

The eligibility criteria for the analysis of AS and AR/RD are shown in Figure 1. For the AS group, CT data were obtained for 60 consecutive patients diagnosed with severe AS indicated for TAVR between October 2020 and October 2021; 2 patients were excluded for anatomical reasons and 8 were excluded due to CT data unavailability or other conditions. The data of the remaining 50 patients were the basis for this analysis. For the AR/RD analysis, 213 patients diagnosed with AR (n=79), root aneurysm (n=32), and patients with both conditions (n=102) underwent aortic valve/root replacement between August 2012 and September 2024. From this cohort, 26 cases were excluded for anatomic reasons, such as bicuspid aortic valve, acute/chronic type-A dissection, and history of valve/root replacement, and 97 cases were excluded due to CT data unavailability or other conditions. Among the remaining 90 cases, 50 were classified into the AR/RD study group, and the other 40 cases with RD were the retraining group. The study group included 38 cases (76%) with mild or more AR and 45 cases (90%) with RD. The etiologies of AR were type I (RD) in 33 cases, type II (leaflet prolapse) in 10 cases, and type III (leaflet restriction) in 8 cases.²² More than 1 etiology type was found in 13 cases (34%). In the retraining group, end-systolic and end-diastolic CT data were individually available in 27 cases, and only end-diastolic data were available in 13 cases, resulting in a total of 67 datasets.

Figure 1.

Eligibility criteria for the (A) AS and (B) AR/RD groups. AS, aortic stenosis; AR, aortic regurgitation; MVR, mitral valve replacement; RD, root dilatation; s/p, status post; TAAD, Type-A aortic dissection; TAVR, transcatheter aortic valve replacement.

Automatic Segmentation and Measurement of the Aortic Valve Leaflets and Root

The AVRM algorithm, developed with 258 CT volume data²⁰ and retrained for RD with the additional 67 CT data, was installed for research purposes on the Vitrea workstation²¹ and used for the analysis. The measured parameters are shown in Figure 2 as referred to a previous study¹³ and defined as follows: the virtual basal ring (VBR) area and perimeter were calculated from the contour of the aortic root region on the nadir plane connecting the 3 nadir points; the sinus of Valsalva (SOV) diameter was the largest distance between each sinus and the opposite commissure side on the cutting plain that made the largest area of the SOV between the nadir plane and its parallel plane on ≥1 commissure; the sinotubular junction (STJ) diameter was the average of the long-axis and short-axis diameters of the smallest aortic contour between the commissure plane (onto all 3 commissures) and the parallel plane located 1 cm above it; the coronary artery height was a straight perpendicular line distance from the coronary ostium point to the nadir plane; gH was the shortest curved line length along the surface of the valve leaflet from a 3-leaflet coaptation point or a midportion of the leaflet’s free margin to a hinge point of the leaflet; eH was the length of a straight perpendicular line from that point of the leaflet’s free margin to a nadir plane.

Figure 2.

Schematic illustration of the measurement parameters of the aortic valve/root. (A) Three-dimensional view. (B) Axial section at the level of the coronary orifices. (C) Sagittal section along the blue dotted line in (B). Measurement parameters are shown in red. eH, effective height; gH, geometric height; RD, root dilatation; SOV, sinus of Valsalva; STJ, sinotubular junction; VBR, virtual basal ring.

Manual Measurement of the Aortic Valves

For the manual measurement of the aortic valve/root, 1 cardiac surgeon (Expert 1, HY), 1 cardiologist (Expert 2, HT), and 1 radiology technician (Expert 3, KI), who had 8-, 6-, and 6-year experience in cardiac CT data analyses, respectively, participated. Expert 1 measured 100 data sets, comprising 50 AS cases and 50 AR/RD cases, whereas Experts 2 and 3 measured 20 of 50 cases in each group. On the multiplanar 2-dimensional reconstruction images of each patient’s CT data using the image processing system (Vitrea), the aortic root structure was manually searched and detected, followed by measurements in a blinded manner as follows: the VBR area and perimeter, left and right coronary artery orifice heights, gH and eH of the left, right, and noncoronary leaflets, the SOV diameter, and the STJ diameter.

Processing and Measurement Time for Aortic Valve/Root Analysis

The processing times for the AVRM and manual measurements were automatically recorded in the Vitrea workstation.

Statistical Analysis

Continuous data are presented as means with standard deviations unless otherwise specified. Differences were calculated between 2 measurements of the same item. Absolute errors (AE) were determined as the absolute values of these differences, and error rates (ER) were calculated as the percentage of AE relative to the measurements. The AE and ER are presented as means with standard errors (SE). The consistency between 2 respective measurements was analyzed using the intraclass correlation coefficient (ICC) with 95% confidence interval (CI). Bland-Altman plots with limits of agreement²³ were analyzed. The fixed error between the algorithm and Expert 1 was analyzed using the paired t-test on their differences. The proportional error was assessed using linear regression analysis between the mean and difference of the algorithm and Expert 1. To compare the total processing time for 40 cases between the algorithm and manual measurements by 3 experts, Dunnett’s test was performed. All statistical analyses were conducted using the Statistical Package for the Social Sciences (version 21; IBM Corp., Armonk, NY, USA). The Bonferroni correction was applied to adjust for multiple validation endpoints across the eight measurement items of the aortic root/valve, with differences considered statistically significant if P<0.0063.

Results

Automated Segmentation and 3D Visualization of the Aortic Valve/Root

The retrained AVRM algorithm successfully visualized 3D shapes of the aortic valve/root in all 100 cases (Figure 3, Supplementary Movie). In 1 case of AS it showed the valve leaflets with a detailed calcification distribution, part of which extended toward the left ventricular outflow tract; another case of AR showed disproportionate RD, leaflet elongation, and uneven commissure heights. The aortic valve leaflets were then selectively visualized to represent gH and eH (Figure 4A: the automated measurements of each parameter are indicated by yellow lines). Even with a bulky calcified region, each parameter was automatically drawn along the anatomic contour as if the valve/root was not calcified. Representative cases of AR with leaflet prolapse (type II) and restriction (type III) without RD demonstrated detailed leaflet shapes (Figure 4B).

Figure 3.

Automated visualization of the aortic valve/root geometry. Three-dimensional shapes of the aortic valve/root in cases of AS (A) and AR/RD (B). Each panel displays: (a) aortic side, (b) left ventricular side, (c) left-coronary leaflet, (d) right-coronary leaflet, (e) noncoronary leaflet, (f–h) VBR, (i,j) SOV diameter on the left-coronary side, (k,l) right-coronary artery height, and (m–o) STJ. Valve leaflets are represented by blue (left-coronary), yellow (right-coronary), and green (noncoronary). The contours of the aortic wall and distances automatically measured by the algorithm are shown by red curves and yellow solid lines, respectively. Calcification is depicted in white. Two-dimensional (2D) images show multiplanar resection (MPR) views corresponding to the respective parameters. The blue, green, and red perpendicular lines in the 2D images represent the cross-sections of the MPR. Only white dotted lines indicate the Expert’s manual measurements for the SOV (B-i) and STJ (B-o) in the AR/RD panel. AS, aortic stenosis; AR, aortic regurgitation; RD, root dilatation; SOV, sinus of Valsalva; STJ, sinotubular junction; VBR, virtual basal ring.

Figure 4.

Automated visualization and measurement of aortic valve leaflets. (A) Three-dimensional shapes of aortic valve leaflets selectively shown (a–c, g–i) with geometric and effective heights indicated by yellow lines in AS and AR/RD. Two-dimensional MPR views corresponding to each measurement are also shown (d–f, j–l). Each panel shows: (a,d,g,j) left-coronary leaflets in blue; (b,e,h,k) right-coronary leaflets in yellow; and (c,f,i,l) noncoronary leaflets in green. (B) Representative aortic valve/root images of leaflet prolapse (type II, a–d) and restriction (type III, e–h) as causes of AR without RD. Automated (c,g) and manual (d,h) tracings for geometric height measurement are shown. AS, aortic stenosis; AR, aortic regurgitation; RD, root dilatation.

Comparison of the Aortic Valve/Root Measurements by the AVRM Algorithm and Expert 1

Using the CT data of 50 cases per group, the measurements of the aortic valve/root geometry by the algorithm and Expert 1 are presented in Table 1. The VBR area and perimeter was 19.4–30.0 mm² (4.6–5.4%) and 1.6–2.2 mm (2.2–2.6%) of AE, respectively, indicating excellent ICC (0.95–0.96). The SOV diameter was 1.3 mm (4.1%) and 2.9 mm (6.5%) of the AE in the AS and AR/RD groups, corresponding to 0.90 and 0.86 of ICC, respectively. The errors slightly increased in the AR/RD group because the dissection plane for measuring the SOV diameter was discrepant between the algorithm and experts in RD (Figure 3B-i). The individual sinus diameters are presented in Supplementary Table 1. The noncoronary sinus tended to be larger than the other sinuses despite no statistical significance. The STJ diameter in the AS group was 2.1 mm (7.3%) of the AE with substantial ICC (0.77) because the STJ waist appeared definitive. However, the AE of the STJ diameter in the AR/RD group increased to 10.4 mm (29.3%) with a low ICC (0.27) because of the indefinite borderline over the dilated sinuses (Figure 3B-o). Expert 1 then remeasured the “STJ diameter” as ordered to the algorithm, resulting in improved AE (3.0 mm, 7.6%) and ICC (0.92). Coronary artery heights were 1.0–1.3 mm (6.2–8.9%) of AE, corresponding to high–excellent ICC (0.81–0.96). Regarding valve leaflet geometry, gH was 1.2–1.8 mm (8.2–10.0%) of the AE with substantial ICC (0.65–0.68), and eH showed 1.3–1.6 mm (12.3–14.6%) of the AE with moderate−substantial ICC (0.55–0.77). In the cases of type II/III AR there was a larger ER compared with type I or no AR in the gH (14.2% vs. 8.0%) and eH (22.7% vs. 10.8%) measurements, with worse ICC (0.57 vs. 0.65) for gH and similar ICC (0.74 vs. 0.74) for eH (Supplementary Table 2).

Table 1.

Measurement Values of Aortic Valve/Root Geometry Using the AVRM Algorithm and Manually by Expert 1

Parameter	Group	Case, n	Algorithm [mm² or mm]	Expert 1 [mm² or mm]	Difference [mm² or mm]	AE (SE) [mm² or mm]	ER (SE) [%]	ICC (95% CI)
VBR area	AS	50	442.77±94.21	448.89±81.20	−6.13±24.33	19.38 (2.22)	4.57 (0.56)	0.96 (0.93~0.98)
VBR area	AR/RD	50	583.81±145.09	597.20±134.21	−13.40±44.17	29.96 (4.94)	5.35 (0.92)	0.95 (0.91~0.97)
VBR perimeter	AS	50	75.59±8.15	74.93±6.98	0.66±1.98	1.64 (0.18)	2.21 (0.24)	0.96 (0.93~0.98)
VBR perimeter	AR/RD	50	87.67±10.17	87.05±9.48	0.62±2.88	2.20 (0.27)	2.56 (0.33)	0.96 (0.92~0.98)
SOV diameter	AS	150	32.23±3.39	31.18±3.38	1.05±1.19	1.25 (0.08)	4.10 (0.27)	0.90 (0.57~0.96)
SOV diameter	AR/RD	150	48.29±7.90	45.66±7.35	2.64±3.20	2.92 (0.24)	6.52 (0.54)	0.86 (0.54~0.94)
STJ diameter	AS	50	29.67±3.47	27.86±2.89	2.07±1.61	2.07 (0.23)	7.25 (0.79)	0.77 (0.05~0.92)
STJ diameter	AR/RD	50	47.65±8.70	37.35±6.55	10.30±7.57	10.39 (1.05)	29.33 (3.08)	0.27 (−0.10~0.59)
STJ diameter within 1 cm above commissure	AR/RD	50	47.65±8.70	45.15±10.08	2.50±2.92	3.04 (0.33)	7.63 (0.99)	0.92 (0.65~0.97)
Right-coronary height	AS	50	16.66±2.97	16.14±3.10	0.52±1.57	1.28 (0.15)	8.26 (0.98)	0.86 (0.75~0.92)
Right-coronary height	AR/RD	50	23.13±6.63	23.26±6.45	−0.13±2.16	1.27 (0.25)	6.19 (1.26)	0.95 (0.91~0.97)
Left-coronary height	AS	50	13.20±2.29	12.66±2.38	0.55±1.36	1.03 (0.15)	8.89 (1.47)	0.81 (0.67~0.89)
Left-coronary height	AR/RD	50	18.73±5.20	18.22±5.09	0.51±1.40	1.17 (0.13)	7.04 (0.85)	0.96 (0.92~0.98)
Geometric height	AS	150	15.68±1.90	15.44±2.07	0.24±1.59	1.23 (0.08)	8.20 (0.59)	0.68 (0.58~0.75)
Geometric height	AR/RD	150	19.08±2.66	18.54±3.09	0.54±2.38	1.76 (0.14)	9.95 (0.84)	0.65 (0.54~0.74)
Effective height	AS	150	12.25±1.55	11.34±1.66	0.92±1.38	1.31 (0.08)	12.34 (0.85)	0.55 (0.24~0.72)
Effective height	AR/RD	150	13.25±2.75	12.36±3.07	0.89±1.84	1.57 (0.11)	14.61 (1.29)	0.77 (0.61~0.85)

AE, absolute error; AR/RD, aortic regurgitation/root dilatation; AS, aortic stenosis; AVRM, aortic valve leaflet and root segmentation and measurement; CI, confidence interval; ER, error rate; ICC, intraclass correlation coefficient; SE, standard error; SOV, sinus of Valsalva; STJ, sinotubular junction; VBR, virtual basal ring.

The Blant-Altman plots of all measurements obtained by the algorithm and Expert 1 are shown in Figure 5. Satisfactory agreement was achieved for measuring VBR dimensions, coronary height, gH, and eH. A proportional error was observed in the VBR dimensions of the AS group. Fixed errors were observed in the left-coronary height and eH in the AS group, and in the eH in the AR/RD group. The SOV and STJ diameters in the AR/RD group showed remarkable dispersion, as compared with the AS group, with significant fixed errors.

Figure 5.

Bland-Altman plots of the automated and manual measurement values by the AVRM algorithm and Expert 1, respectively, for the VBR area and perimeter (A,B), SOV diameters (C), STJ diameter (D), right and left-coronary heights (E,F), and geometric and effective heights (G,H). The round orange markers and triangular green markers represent the AS and AR/RD groups, respectively. The red and green solid lines represent the mean differences between the algorithm and Expert 1 in the AS and AR/RD groups, respectively. The red and green dashed lines show the limits of agreements (±1.96 standard deviation) in the AS and AR groups, respectively. N=50 for VBR area and perimeter, STJ diameter, right and left-coronary heights per group. N=150 for SOV diameter, geometric height, and effective height per group by including 3 aortic valve leaflets per case. AS, aortic stenosis; AR, aortic regurgitation; AVRM, aortic valve leaflet and root segmentation and measurement; RD, root dilatation; SOV, sinus of Valsalva; STJ, sinotubular junction; VBR, virtual basal ring.

Assessment of the Correlation of Measurement Values Among the AVRM Algorithm and 3 Experts

The CT data of 20 cases per group were analyzed by Experts 2 and 3, and the correlation of measurement values among the algorithm and the 3 experts was assessed. In the AS group (Table 2), excellent ICC in the VBR dimension, SOV, and STJ diameter; high ICC in the coronary height; and moderate ICC in the gH and eH were observed between the experts, but the comparison between the algorithm and Expert 1 showed similar ICC in the VBR dimension (0.96–0.97 vs. 0.94–0.97) and eH (0.50 vs. 0.43–0.51); decreased ICCs in SOV diameter (0.89 vs. 0.94), STJ diameter (0.81 vs. 0.98), and coronary height (0.69–0.78 vs. 0.80–0.90); and rather increased ICC (0.66 vs. 0.48) in the gH analysis. In the AR/RD group (Table 3), excellent ICC in the VBR dimension and coronary height; substantial–high ICC in SOV and STJ diameter, gH, and eH, were observed between the experts. In contrast, comparison between the algorithm and Expert 1 showed similar ICC in the VBR dimension (0.92–0.95 vs. 0.90–0.96), SOV diameter (0.87 vs. 0.82–0.86), coronary height (0.91–0.96 vs. 0.92–0.98), and decreased ICC in STJ diameter (0.32 vs. 0.77–0.90), gH (0.66 vs. 0.79–0.87), and eH (0.52 vs. 0.83–0.87). Of note, a comparison between the mean values of the experts and the algorithm revealed better ER and ICC in the VBR dimension, SOV diameter, and eH in the AS and AR/RD groups and gH in the AS group, suggesting consensus measurement values of the experts may better approximate the algorithm’s results by cancelling each expert’s errors.

Table 2.

Assessment of the Correlation of Measurement Values Among the AVRM Algorithm and 3 Experts in 20 AS Cases

Parameter	Comparison	Difference [mm² or mm]	AE (SE) [mm² or mm]	ER (SE) [%]	ICC (95% CI)
VBR area (n=20)	Expert 1 vs. Algorithm	−5.92±27.41	20.81 (4.08)	4.74 (0.95)	0.97 (0.89~0.99)
	vs. Expert 2	18.01±23.19	21.36 (4.47)	5.20 (1.14)	0.94 (0.74~0.98)
	vs. Expert 3	7.48±28.26	24.34 (3.42)	5.40 (0.78)	0.95 (0.88~0.98)
	Expert (mean) vs. Algorithm	−2.57±19.11	15.47 (2.45)	3.55 (0.56)	0.98 (0.95~0.99)
VBR perimeter (n=20)	Expert 1 vs. Algorithm	−0.79±1.99	1.65 (0.30)	2.15 (0.37)	0.96 (0.63~0.99)
	vs. Expert 2	0.83±1.73	1.25 (0.32)	1.72 (0.46)	0.97 (0.90~0.99)
	vs. Expert 3	−1.28±2.23	2.19 (0.29)	2.84 (0.36)	0.94 (0.46~0.98)
	Expert (mean) vs. Algorithm	−0.64±1.30	1.26 (0.15)	1.67 (0.19)	0.98 (0.95~0.99)
SOV diameter (n=60)	Expert 1 vs. Algorithm	−2.00±0.93	2.00 (0.12)	6.69 (0.43)	0.89 (0.09~0.97)
	vs. Expert 2	−0.65±0.93	0.85 (0.10)	2.77 (0.32)	0.94 (0.82~0.98)
	vs. Expert 3	−0.21±1.17	0.93 (0.09)	3.10 (0.32)	0.94 (0.88~0.96)
	Expert (mean) vs. Algorithm	−1.71±0.76	1.71 (0.10)	5.65 (0.33)	0.86 (−0.03~0.97)
STJ diameter (n=20)	Expert 1 vs. Algorithm	−1.80±0.99	1.80 (0.22)	6.43 (0.74)	0.81 (−0.02~0.95)
	vs. Expert 2	0.01±0.56	0.47 (0.06)	1.66 (0.20)	0.98 (0.96~0.99)
	vs. Expert 3	−0.26±0.53	0.46 (0.08)	1.64 (0.30)	0.98 (0.94~0.99)
	Expert (mean) vs. Algorithm	−1.72±1.08	1.72 (0.24)	6.12 (0.80)	0.83 (−0.02~0.96)
Right-coronary height (n=20)	Expert 1 vs. Algorithm	−1.21±1.06	1.33 (0.20)	9.61 (1.67)	0.69 (−0.08~0.92)
	vs. Expert 2	0.66±1.35	1.03 (0.24)	7.65 (1.85)	0.80 (0.53~0.92)
	vs. Expert 3	0.47±1.26	1.12 (0.16)	8.02 (1.26)	0.86 (0.69~0.94)
	Expert (mean) vs. Algorithm	−1.58±0.58	1.58 (0.13)	11.48 (0.96)	0.79 (−0.04~0.95)
Left-coronary height (n=20)	Expert 1 vs. Algorithm	−0.79±0.85	0.96 (0.15)	8.09 (1.31)	0.78 (0.46~0.91)
	vs. Expert 2	−0.04±1.00	0.81 (0.13)	7.33 (1.38)	0.90 (0.77~0.96)
	vs. Expert 3	0.21±1.22	0.96 (0.17)	8.53 (1.84)	0.83 (0.63~0.93)
	Expert (mean) vs. Algorithm	−0.85±0.96	1.08 (0.15)	9.33 (1.41)	0.83 (0.35~0.94)
Geometric height (n=60)	Expert 1 vs. Algorithm	−0.44±1.86	1.50 (0.15)	10.23 (1.08)	0.66 (0.41~0.80)
	vs. Expert 2	0.42±2.25	1.82 (0.18)	12.72 (1.30)	0.48 (0.26~0.65)
	vs. Expert 3	−1.05±2.05	1.76 (0.19)	12.01 (1.39)	0.48 (0.16~0.69)
	Expert (mean) vs. Algorithm	−0.22±1.13	0.96 (0.08)	6.32 (0.52)	0.84 (0.75~0.90)
Effective height (n=60)	Expert 1 vs. Algorithm	−1.41±1.51	1.77 (0.13)	17.46 (1.43)	0.50 (0.29~0.67)
	vs. Expert 2	−0.93±1.81	1.51 (0.17)	12.46 (1.27)	0.43 (0.17~0.62)
	vs. Expert 3	−0.31±1.43	1.03 (0.13)	9.88 (1.24)	0.51 (0.29~0.67)
	Expert (mean) vs. Algorithm	−1.00±1.35	1.32 (0.13)	12.59 (1.34)	0.58 (0.19~0.78)

Abbreviations as in Table 1.

Table 3.

Assessment of the Correlation of Measurement Values Among the AVRM Algorithm and 3 Experts in 20 AR/RD Cases

Parameter	Comparison	Difference [mm² or mm]	AE (SE) [mm² or mm]	ER (SE) [%]	ICC (95% CI)
VBR area (n=20)	Expert 1 vs. Algorithm	22.62±44.57	33.97 (8.09)	5.87 (1.30)	0.95 (0.88~0.98)
	vs. Expert 2	31.80±27.39	32.22 (6.01)	5.87 (0.97)	0.90 (0.28~0.97)
	vs. Expert 3	15.43±28.68	21.05 (5.51)	3.50 (0.73)	0.96 (0.73~0.99)
	Expert (mean) vs. Algorithm	6.88±31.39	22.02 (5.12)	3.98 (0.98)	0.94 (0.85~0.97)
VBR perimeter (n=20)	Expert 1 vs. Algorithm	−0.06±2.98	2.33 (0.40)	2.69 (0.43)	0.92 (0.76~0.97)
	vs. Expert 2	1.26±2.10	1.84 (0.36)	2.15 (0.39)	0.93 (0.79~0.98)
	vs. Expert 3	−0.17±2.60	1.64 (0.44)	1.88 (0.49)	0.95 (0.63~0.99)
	Expert (mean) vs. Algorithm	−0.42±2.30	1.91 (0.28)	2.25 (0.35)	0.94 (0.86~0.98)
SOV diameter (n=60)	Expert 1 vs. Algorithm	−5.09±3.10	5.09 (0.40)	11.59 (0.91)	0.87 (0.31~0.96)
	vs. Expert 2	−2.57±2.09	2.72 (0.24)	5.79 (0.50)	0.86 (0.18~0.96)
	vs. Expert 3	0.45±1.36	1.04 (0.13)	2.23 (0.25)	0.82 (0.05~0.94)
	Expert (mean) vs. Algorithm	−4.38±2.59	4.38 (0.33)	9.67 (0.70)	0.75 (−0.06~0.92)
STJ diameter(n=20)	Expert 1 vs. Algorithm	−8.10±7.43	8.31 (1.60)	21.29 (4.17)	0.32 (−0.08~0.64)
	vs. Expert 2	−0.90±5.09	2.32 (1.03)	4.56 (1.66)	0.77 (0.51~0.90)
	vs. Expert 3	0.84±3.14	2.37 (0.48)	6.15 (1.25)	0.90 (0.74~0.96)
	Expert (mean) vs. Algorithm	−8.08±8.24	8.65 (1.70)	23.07 (4.81)	0.31 (−0.09~0.64)
Right-coronary height (n=20)	Expert 1 vs. Algorithm	−0.71±1.27	1.12 (0.20)	6.16 (1.42)	0.96 (0.48~0.99)
	vs. Expert 2	0.66±1.22	1.15 (0.17)	6.20 (1.00)	0.97 (0.92~0.99)
	vs. Expert 3	0.79±0.99	1.07 (0.15)	5.76 (0.92)	0.98 (0.96~0.99)
	Expert (mean) vs. Algorithm	−1.19±0.88	1.31 (0.15)	7.06 (0.99)	0.97 (0.53~0.99)
Left-coronary height (n=20)	Expert 1 vs. Algorithm	−1.05±1.04	1.21 (0.18)	8.57 (1.38)	0.91 (0.55~0.97)
	vs. Expert 2	−0.08±0.63	0.49 (0.09)	3.33 (0.63)	0.98 (0.96~0.99)
	vs. Expert 3	0.84±0.96	1.07 (0.15)	7.09 (1.05)	0.92 (0.62~0.97)
	Expert (mean) vs. Algorithm	−1.30±0.99	1.35 (0.20)	9.77 (1.58)	0.88 (0.14~0.97)
Geometric height (n=60)	Expert 1 vs. Algorithm	−1.25±2.08	1.90 (0.19)	10.85 (1.24)	0.66 (0.15~0.84)
	vs. Expert 2	0.33±9.71	0.92 (0.11)	5.34 (0.74)	0.87 (0.79~0.92)
	vs. Expert 3	1.15±1.71	1.52 (0.18)	8.37 (0.92)	0.79 (0.59~0.89)
	Expert (mean) vs. Algorithm	−1.75±1.75	2.05 (0.18)	12.13 (1.25)	0.61 (0.06~0.83)
Effective height (n=60)	Expert 1 vs. Algorithm	−1.67±1.72	1.98 (0.17)	18.85 (2.33)	0.52 (0.11~0.74)
	vs. Expert 2	−0.25±1.16	0.82 (0.11)	6.92 (0.96)	0.87 (0.79~0.92)
	vs. Expert 3	0.08±1.01	0.75 (0.09)	6.43 (0.78)	0.83 (0.73~0.89)
	Expert (mean) vs. Algorithm	−1.61±1.54	1.85 (0.16)	16.94 (1.94)	0.55 (0.00~0.79)

Abbreviations as in Table 1.

Processing and Measurement Times

The manual measurement times of all components per case were 618±91 s, 672±180 s, and 1,126±468 s by Experts 1, 2, and 3, respectively (n=40). Using the same cohort, the total processing time for the AVRM per case was 122±54 s, which was significantly shorter than the experts’ measurement times (P<0.001).

Discussion

Automating the aortic root measurement from CT data could standardize treatment planning for aortic valve/root diseases, and several studies have reported deep learning-based methods for root segmentation in TAVR planning.¹⁴^–¹⁹ Santaló-Corcoy et al.¹⁹ proposed a fully automated algorithm known as “TAVI-PREP” and validated it with expert manual measurements in 200 cases. They demonstrated a high correlation (AE, 5%) between the algorithm and the experts for VBR, left ventricular outflow tract, STJ, and SOV dimensions, except for coronary height (AE, 11.6–16.5%). Saitta et al. reported comparable accuracy of their algorithm for the VBR and STJ.¹⁸ Astudillo et al. showed better accuracy than TAVI-PREP for coronary height.¹⁶ Although these algorithms were focused on the essential requirement for TAVR, detailed information on the valve leaflets is missing. For a safe TAVR procedure, it is important to evaluate valve/root calcification to avoid root rupture and coronary obstruction. Rouhollahi et al. proposed an algorithm to detect and quantify aortic valve calcification and projected it over patient-oriented leaflet shapes.²⁴ However, the leaflet was made using a virtual surface mesh interpolating several leaflet landmarks and not reflecting the details of their shapes.

Basically, visualizing the highly mobile, thin valve leaflets in a cardiac cycle is challenging. Some previous studies have proposed deep learning-based algorithms to segment aortic valve leaflets from CT images, but their systems were not fully automated. Some of them required manual data cropping around the valve,²⁵ while others omit parts of the aortic root in their segmentation.²⁶ To address these challenges, we developed the fully automated AVRM algorithm using CT images,²⁰ combining 2 deep neural networks (Spatial Configuration-Net²⁷ for detecting anatomical landmarks and U-Net²⁸ for segmenting aortic valve leaflets) and achieving high accuracy in the segmentation process.

To the best of our knowledge, this study is the first to compare a fully automated algorithm’s measurements of aortic valve leaflets with experts’ manual measurements. Our updated algorithm accurately visualized the patient-specific 3D aortic valve/root morphology with detailed valve/root calcification. This would help sharing of the overall images and planning both surgical and transcatheter interventions. The AVRM showed excellent precision of VBR measurement in the AS group with the difference of 6.1 mm² (area) and 0.7 mm (perimeter), which were even better than the TAVI-PREP (9.7 mm² and 0.7 mm)¹⁹ and Saitta et al.’s algorithm (perimeter, 1.8 mm).¹⁸ The VBR dimension is the most critical determinant of TAVR device selection. For example, the balloon expandable SAPIEN 3 (Edwards Lifesciences, CA, USA) has 4 sizes (20, 23, 26, and 29 mm) depending on the VBR area.²^,²⁹ The area range for shifting the SAPIEN size is 72–143 mm², indicating that the AE (19.4 mm²) observed in this study is small enough. The self-expandable Evolut^TM FX (Medtronic plc, MN, USA) has 4 sizes (23, 26, 29, and 34 mm) determined by the VBR perimeter.³^,³⁰ Considering the perimeter range to change the Evolut size is 6.3–12.5 mm, the AE (1.6 mm) in our study can be justified. The precise algorithm for VBR sizing in the AR/RD cases would also be useful for surgical valve/root repair, ensuring the selection of appropriate graft and annular ring sizes.

Our algorithm gave slightly larger SOV and STJ diameters than the experts’ measurements for AS, with ER of 4.1% and 7.3%, respectively. Similar results were reported for TAVI-PREP, indicating an ER of 3–5% for the SOV and 2–6% for the STJ.¹⁹ For AR/RD, our algorithm measured SOV diameters 6.5% larger than Expert 1, presumably because the algorithm was ordered to seek the largest SOV from the VBR up to “≥1 commissure,” whereas the experts sought the largest SOV between the VBR and the plane “covering all commissures” (Figure 3B-i). Similarly, the algorithm was ordered to seek the STJ as the narrowest aortic portion between the commissure plane and 1 cm above it. However, the experts misunderstood this “STJ definition” and found the narrowest part far beyond the commissures (Figure 3B-o). Thus, our algorithm successfully segmented all cases and measured these parameters faithfully as ordered. The main reason for inconsistent SOV and STJ measurements was mismatch of the parameters’ definitions.

Coronary height is another crucial factor in aortic valve treatments for predicting the risk of coronary obstruction. This study showed a satisfactory correlation (ICC, 0.81–0.86) between the AVRM algorithm and Expert 1 in the AS group compared with TAVI-PREP (correlation coefficient, 0.72–0.80)¹⁹ and Astudillo et al.’s algorithm (0.80).¹⁶

Calcified bulk over the aortic annulus, leaflets, and STJ can hinder accurate measurement of valve/root dimensions.¹⁹ The AVRM algorithm successfully traced the VBR, STJ, and gH even under heavy calcification, as it precisely detected target points and lines in relation to adjacent non-calcified landmarks in the 3D mapping of the whole valve structure even if they were involved in the calcified bulk (Figures 3A,4A). In contrast, as the experts measured the leaflets on the bisected 2D plane, it was more challenging to draw the line for gH measurement on the heavily calcified leaflets without 3D mapping, resulting in unconfirmed line detection at the center or surface of the thickened calcified leaflets. This fuzziness likely led to the large ER (12.0–12.7%) among the experts. On the other hand, improved ER of the gH measurement was observed between the experts for non-calcified leaflets in the AR/RD case (5.3–8.4%).

Precise information on respective sinuses is worth having because a small SOV can be a risk for coronary obstruction during TAVR. The sizes of the 3 leaflet-related sinuses are uneven, with a large noncoronary sinus and a small right-coronary sinus.¹³ Our results were similar despite no statistical significance (Supplementary Table 1). Additionally, gH and eH correlate with SOV size to ensure proper leaflet coaptation,¹³^,³¹ with the mean values of gH and eH in normal Japanese hearts being 14.7±1.3 mm and 8.6±1.4 mm, respectively.¹³ Our results showed a large gH in the noncoronary leaflets and a small gH in the right-coronary leaflets, whereas eH was even among the leaflets. Compared with normal values, our results seemed reasonable considering the etiologies of AS and AR/RD.

This nonuniformity in valve/root geometry must be considered during valvuloplasty and VSRR as well as TAVR. Surgeons typically assess the valve under cardiac arrest in a short time frame, and their experience is critical for the procedure’s success.⁶ The AVRM algorithm offers an objective, reproducible, and precise assessment of root/valve geometry, which can be shared with the surgical team to assist preoperative procedural planning and intraoperative decision-making. Moreover, our algorithm reduces the clinical workload by shortening measurement times. Even compared with Expert 1, the fastest among the experts, the algorithm saved approximately 500 s/case. Notably, the algorithm’s processing time remained consistent regardless of the number of measured parameters, making it efficient for scaling to more complex cases.

Study Limitations

First, cases of bicuspid aortic valves, type-A dissection, and history of valve/root replacement were excluded. Although our study group included 16 cases (32%) of prolapsed and/or restricted aortic valves and showed fine 3D visualization, the majority of cases had RD and in the cases of leaflet deformities the algorithm was less precise (Supplementary Table 2). Our algorithm requires further retraining for various aortic valve/root deformities. Second, the sample size for retraining the RD was relatively small. A larger sample size is expected to further improve the algorithm’s accuracy. Third, because most of the cases in this study used CT data from a single center, the number and distribution of cases may not have been sufficient to fully validate the algorithm. Fourth, the algorithm requires adjusted SOV and STJ definitions to better align with clinical needs. Finally, the reconstructed aortic valve leaflets may be thicker than the real structure because the algorithm detects the leaflet surfaces as contrast boundaries.²⁵ Consequently, reconstructing fenestrations or string-like structures thinner than 1 mm remains challenging. Improvements in the algorithm and CT data acquisition are expected to lead to more accurate reconstruction of aortic valve fine structures.

Conclusions

Fully automated measurement of aortic valve/root geometry from ECG-gated CT data was successfully achieved using our updated deep learning-based algorithm for RD. The algorithm provided patient-specific 3D aortic valve/root visualization. It demonstrated satisfactory consistency in the measurement of the VBR, SOV, coronary height, and STJ against standard manual measurements for AS, surpassing recent TAVR-oriented algorithms by offering additional detailed leaflet measurements. For AR/RD, the proposed algorithm also provided accurate VBR, coronary height, and leaflet information, potentially useful for valve-sparing surgery. However, there is still room for improvement of the current algorithm in measuring the SOV and STJ dimensions upon RD. This deep learning-based approach can assist physicians in evaluating aortic valve/root anatomy and planning treatment strategies while minimizing workload and saving time.

Acknowledgments

This study was funded by the Center of Innovation Program from the Japan Science and Technology Agency, JST, under Grant Number JPMJCE1304, and the Canon Medical Systems Corporation. The authors are deeply grateful to Dr. Eriko Maeda and Dr. Asuka Hatano for their tremendous contribution to this project. We also thank Enago (https://www.enago.com/) for English language editing.

Disclosures

I.S. received research grants from Canon Medical Systems Corporation. G.A., T.S., and K.F. are Canon Medical Systems Corporation employees. M.O. is a member of Circulation Journal’s Editorial Team. The other authors have no conflicts of interest to declare.

IRB Information

This study was approved by the Research Ethics Committee of the Faculty of Medicine, The University of Tokyo, on January 27, 2017 (approval number: 11330).

Data Availability

The deidentified participant data will not be shared.

Supplementary Files

Supplementary Movie. Demonstration of fully automatic aortic valve/root segmentation and measurement from computed tomography data for patients with aortic stenosis and aortic root dilatation.

Please find supplementary file(s);

https://doi.org/10.1253/circj.CJ-24-1031

References

1. Thaden JJ, Nkomo VT, Enriquez-Sarano M. The global burden of aortic stenosis. Prog Cardiovasc Dis 2014; 56: 565–571.
2. Mack MJ, Leon MB, Thourani VH, Makkar R, Kodali SK, Russo M, et al.; PARTNER 3 Investigators. Transcatheter aortic-valve replacement with a balloon-expandable valve in low-risk patients. N Engl J Med 2019; 380: 1695–1705.
3. Popma JJ, Deeb GM, Yakubov SJ, Mumtaz M, Gada H, O’Hair D, et al.; Evolut Low Risk Trial Investigators. Transcatheter aortic-valve replacement with a self-expanding valve in low-risk patients. N Engl J Med 2019; 380: 1706–1715.
4. Sugiyama Y, Moriyama N, Miyashita H, Yokoyama H, Ochiai T, Shishido K, et al. Long-term assessment of survival after transcatheter aortic valve implantation: Insights from the international transcatheter aortic valve implantation registry. Circ J 2024; 88: 462–471.
5. Yokohama F, Takaya Y, Ichikawa K, Nakayama R, Miki T, Toda H, et al. Association between aortic valve calcification and severity of concomitant aortic regurgitation in patients with severe aortic stenosis. Circ J 2024; 88: 606–611.
6. Komiya T. Aortic valve repair update. Gen Thorac Cardiovasc Surg 2015; 63: 309–319.
7. Akinseye OA, Pathak A, Ibebuogu UN. Aortic valve regurgitation: A comprehensive review. Curr Probl Cardiol 2018; 43: 315–334.
8. Stamou SC, Williams ML, Gunn TM, Hagberg RC, Lobdell KW, Kouchoukos NT. Aortic root surgery in the United States: A report from the Society of Thoracic Surgeons database. J Thorac Cardiovasc Surg 2015; 149: 116–122.e4.
9. Arimura S, Seki M, Sasaki K, Takai H, Matsuhama M, Kunihara T, et al. A nationwide survey of aortic valve surgery in Japan: Current status of valve preservation in cases with aortic regurgitation. Gen Thorac Cardiovasc Surg 2017; 65: 429–434.
10. Schäfers HJ, Bierbach B, Aicher D. A new approach to the assessment of aortic cusp geometry. J Thorac Cardiovasc Surg 2006; 132: 436–438.
11. Schäfers HJ, Schmied W, Marom G, Aicher D. Cusp height in aortic valves. J Thorac Cardiovasc Surg 2013; 146: 269–274.
12. Blanke P, Weir-McCall JR, Achenbach S, Delgado V, Hausleiter J, Jilaihawi H, et al. Computed tomography imaging in the context of transcatheter aortic valve implantation (TAVI)/transcatheter aortic valve replacement (TAVR): An Expert Consensus Document of the Society of Cardiovascular Computed Tomography. JACC Cardiovasc Imaging 2019; 12: 1–24.
13. Izawa Y, Mori S, Tretter JT, Quintessenza JA, Toh H, Toba T, et al. Normative aortic valvar measurements in adults using cardiac computed tomography: A potential guide to further sophisticated aortic valve-sparing surgery. Circ J 2021; 85: 1059–1067.
14. Zheng Y, John M, Liao R, Boese J, Kirschstein U, Georgescu B, et al. Automatic aorta segmentation and valve landmark detection in C-arm CT: Application to aortic valve implantation. Springer: Berlin, 2010; 476–483.
15. Elattar MA, Wiegerinck EM, Planken RN, Van Bavel E, Van Assen HC, Baan J Jr, et al. Automatic segmentation of the aortic root in CT angiography of candidate patients for transcatheter aortic valve implantation. Med Biol Eng Comput 2014; 52: 611–618.
16. Astudillo P, Mortier P, Bosmans J, De Backer O, de Jaegere P, Iannaccone F, et al. Automatic detection of the aortic annular plane and coronary ostia from multidetector computed tomography. J Interv Cardiol 2020; 2020: 9843275.
17. Krüger N, Meyer A, Tautz L, Hüllebrand M, Wamala I, Pullig M, et al. Cascaded neural network-based CT image processing for aortic root analysis. Int J Comput Assist Radiol Surg 2022; 17: 507–519.
18. Saitta A, Sturla F, Gorla R, Oliva OA, Votta E, Bedogni F, et al. A CT-based deep learning system for automatic assessment of aortic root morphology for TAVI planning. Comput Biol Med 2023; 163: 107147.
19. Santaló-Corcoy M, Corbin D, Tastet O, Lesage F, Modine T, Asgar A, et al. TAVI-PREP: A deep learning-based tool for automated measurements extraction in TAVI planning. Diagnostics (Basel) 2023; 13: 3181.
20. Aoyama G, Zhao L, Zhao S, Xue X, Zhong Y, Yamauchi H, et al. Automatic aortic valve cusps segmentation from CT images based on the cascading multiple deep neural networks. J Imaging 2022; 8: 11.
21. US Food & Drug Administration. 510(k) premarket notification. https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfPMN/pmn.cfm?ID=K203312 (accessed December 14, 2024).
22. Boodhwani M, de Kerchove L, Glineur D, Poncelet A, Rubay J, Astarci P, et al. Repair-oriented classification of aortic insufficiency: Impact on surgical techniques and clinical outcomes. J Thorac Cardiovasc Surg 2009; 137: 286–294.
23. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986; 1: 307–310.
24. Rouhollahi A, Willi JN, Haltmeier S, Mehrtash A, Straughan R, Javadikasgari H, et al. CardioVision: A fully automated deep learning package for medical image segmentation and reconstruction generating digital twins for patients with aortic stenosis. Comput Med Imaging Graph 2023; 109: 102289.
25. Fan B, Tomii N, Tsukihara H, Maeda E, Yamauchi H, Nawata K, et al. Attention-guided decoder in dilated residual network for accurate aortic valve segmentation in 3D CT scans. In: Liao, H, Balocco S, Wang G, Zhang F, Liu Y, Ding Z, et al, editors. Machine learning and medical engineering for cardiovascular health and intravascular imaging and computer assisted stenting. First International Workshop, MLMECH 2019, and 8^th Joint International Workshop, CVII-STENT 2019, held in conjunction with MICCAI 2019, Shenzen, China, October 13, 2019, Proceedings. Springer: Cham, 2019; 121–129.
26. Pak DH, Caballero A, Sun W, Duncan JS. Efficient aortic valve multilabel segmentation using a spatial transformer network. In: Proceedings of 2020 IEEE 17th international symposium on biomedical imaging (ISBI) Apr 3 2020; 1738–1742. https://ieeexplore.ieee.org/document/9098378 (accessed April 2025).
27. Payer C, Štern D, Bischof H, Urschler M. Integrating spatial configuration into heatmap regression based CNNs for landmark localization. Med Image Anal 2019; 54: 207–219.
28. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 2021; 18: 203–211.
29. Edwards SAPIEN 3 Transcatheter Heart Valve System, SAPIEN 3 Transcatheter Heart Valve, Edwards Commander Delivery System. https://www.accessdata.fda.gov/cdrh_docs/pdf14/P140031S085D.pdf (accessed December 14, 2024).
30. Medtronic Evolut^TM FX System, Evolut^TM FX Transcatheter Aortic Valve, Evolut^TM FX Delivery Catheter System, Evolut^TM FX Loading System. https://www.medtronic.com/content/dam/emanuals/cardio/M004015C001DOC1_RevD_view.pdf (accessed December 14, 2024).
31. Tamer S, Mastrobuoni S, van Dyck M, Navarra E, Bollen X, Poncelet A, et al. Free margin length and geometric height in aortic root dilatation and leaflet prolapse: implications for aortic valve repair surgery. Eur J Cardiothorac Surg 2020; 57: 124–132.

Corresponding author

Register with J-STAGE for free!