A Deep Learning-Based Fully Automated Pipeline for Regurgitant Mitral Valve Anatomy Analysis From 3D Echocardiography

TL;DR

The paper addresses the challenge of time-consuming and operator-dependent mitral valve morphology assessment from 3DTEE by introducing a fully automatic pipeline. It centers on a multi-decoder 3D CNN that jointly segments the mitral annulus, anterior leaflet, and posterior leaflet, followed by a refinement stage and landmark-based morphometry to produce repeatable MV measurements. Key findings show the ensemble CNN achieves a Dice score of $0.82 \pm 0.06$ for the overall mask and a mean surface distance of $0.43 \pm 0.14$ mm pre-refinement, with the pipeline delivering rapid, end-to-end morphology quantification (≈$14.7$ s per volume) and robust agreement with ground-truth measurements as well as favorable comparisons to semi-automated TomTec measurements. The method offers substantially faster and more reproducible MV anatomy analysis, with potential to enhance preprocedural planning and intraoperative guidance for TEER and other mitral interventions, and it opens avenues for broader clinical validation and real-time deployment.

Abstract

Three-dimensional transesophageal echocardiography (3DTEE) is the recommended imaging technique for the assessment of mitral valve (MV) morphology and lesions in case of mitral regurgitation (MR) requiring surgical or transcatheter repair. Such assessment is key to thorough intervention planning and to intraprocedural guidance. However, it requires segmentation from 3DTEE images, which is timeconsuming, operator-dependent, and often merely qualitative. In the present work, a novel workflow to quantify the patient-specific MV geometry from 3DTEE is proposed. The developed approach relies on a 3D multi-decoder residual convolutional neural network (CNN) with a U-Net architecture for multi-class segmentation of MV annulus and leaflets. The CNN was trained and tested on a dataset comprising 55 3DTEE examinations of MR-affected patients. After training, the CNN is embedded into a fully automatic, and hence fully repeatable, pipeline that refines the predicted segmentation, detects MV anatomical landmarks and quantifies MV morphology. The trained 3D CNN achieves an average Dice score of $0.82 \pm 0.06$, mean surface distance of $0.43 \pm 0.14$ mm and 95% Hausdorff Distance (HD) of $3.57 \pm 1.56$ mm before segmentation refinement, outperforming a state-of-the-art baseline residual U-Net architecture, and provides an unprecedented multi-class segmentation of the annulus, anterior and posterior leaflet. The automatic 3D linear morphological measurements of the annulus and leaflets, specifically diameters and lengths, exhibit differences of less than 1.45 mm when compared to ground truth values. These measurements also demonstrate strong overall agreement with analyses conducted by semi-automated commercial software. The whole process requires minimal user interaction and requires approximately 15 seconds

Figures (7)

• Figure 1: Intermediate and final result of the steps followed to manually annotate images. (a) 3D rendering of the segmentation mask for each step, and (b) corresponding cross-sectional view overlapped on the image.
• Figure 2: Schematic representation of the implemented automatic pipeline. SH = saddle horn, PAM = posterior annulus mid-point, MC = medial commissure, LC = lateral commissure, $D_{CC}$ = inter-commissural diameter, $D_{AP}$ = antero-posterior diameter, $L_{A}$ = annulus length, $\bm \Gamma_{A}$= annulus surface, Tips = leaflets tips $L_{PL}$ = posterior leaflet length, $L_{AL}$ = anterior leaflet length, $\bm \Gamma_{AL}$ = anterior leaflet surface, $\bm \Gamma_{PL}$ = posterior leaflet surface.
• Figure 3: Schematic representation of the algorithms for the annulus reconstruction and coaptation line identification. (a) From the top left to the bottom right, the three steps involved in the annulus reconstruction: 1) identification of intersecting points between $\bm \Omega_{A}$ (green) and the multiple rotating planes (gray); 2) interpolation of the intersecting points for the annulus skeleton (yellow) reconstruction; 3) radial expansion of the annulus skeleton. (b) The reference plane $\bm \Pi_{\perp}$ and two representative examples of the multiple rotating planes $\bm \Pi_{\perp}^{30}$ and $\bm \Pi_{\perp}^{-30}$ (gray); the intersecting points (green) between $\bm \Pi_{\perp}^{i}$ and $\bm \Omega\textsubscript{AL}$.
• Figure 4: Illustrative example of reconstructed model with anatomical features and measurements extracted by the automatic pipeline. (a) Reconstructed and refined mitral annulus ($\bm \Omega_{A}$) together with the annular anatomical landmarks (SH = saddle horn, PAM = posterior annulus mid-point, MC = medial commissure, LC = lateral commissure), the best fitting plane ($\bm \Pi$) and the unitary vector (n) normal to $\bm \Pi$. (b) Reconstructed mitral leaflets ($\bm \Omega_{AL}$=anterior leaflet surface, in red; $\bm \Omega_{PL}$=posterior leaflet surface, in blue), main leaflet landmarks (Tips = leaflets tips), and the reconstructed model of the coaptation line (green). (c) Leaflet 3D middle surfaces ($\bm \Gamma_{AL}$, $\bm \Gamma_{PL}$) and leaflet 3D measurements ($L_{PL}$ = posterior leaflet length, $L_{AL}$ = anterior leaflet length). (d) Color-coded 3D representation of the reconstructed representing leaflet height. (e) 3D surface interpolating the non-planar annular profile ($\bm \Gamma_{A}$) and defining the annular area, and annulus 3D measurements ($D_{CC}$ = inter-commissural diameter, $D_{AP}$ = antero-posterior diameter, $L_{A}$ = annulus length, $H_{A}$ = annulus height).
• Figure 5: Training and validation loss curves for each validation subset (f0-f9) over the total amount of epochs. Continuous line represents the training loss while the dashed line represents the validation loss.