Segmentation of Tricuspid Valve Leaflets From Transthoracic 3D Echocardiograms of Children With Hypoplastic Left Heart Syndrome Using Deep Learning

Introduction/ Abstract

Hypoplastic Left Heart Syndrome (HLHS)
HLHS occurs when only the right ventricle and tricuspid valve (TV) support circulation.

• TV → Tricuspid Valve

• 3DE → 3D Echocardiography

• TEE → Transesophageal Echocardiography

• Used an FCN to segment TVs from transthoracic 3DE

• Dataset split: 133 3DE scans for training, 28 for validation

• Metrics used: Dice Similarity Coefficient (DSC) and Mean Boundary Distance (MBD)

• Results:

  • Annular curve
    • DSC (median): 0.86
    • MBD (median): 0.35
  • Merged
    • DSC (average): 0.77
    • MBD (median): 0.66

Anatomical details:

• Annulus: the fibrous ring that serves as the base structure for the valves
• Leaflets: the thin flaps that open and close with each heartbeat — think of these as hinges and doors

Repairing the TV is difficult due to its complex geometry.

Limitations

• 2DE: requires mental reconstruction by the clinician
• 3DE: pediatric transthoracic scans are of lower quality compared to adults
• TEE: affected by artifacts and limited cooperation in children

Dataset Details

Basic Information

• 161 TEE 3DE images from 239 unique HLHS patients
• 133 for training (pre-stage 1: 24; post-stage 1: 21; post-stage 2: 12; post-stage 3: 76)
• 28 for validation/testing

Ground Truth Creation

1. Annular curve is marked
2. TV leaflets are segmented
3. Valve quadrant landmarks corresponding to APSL (Anterior, Posterior, Septal, Lateral) regions of the annulus are identified
4. Commissures (boundaries between the leaflets and annulus) are marked: ASC, PSC, APC

The Annular Curve with quadrant landmarks

Model Architecture

• Base architecture: VNet
• Modifications:
  • Changed activation from PReLU to ReLU based on experimental feedback
  • Changed convolution filter sizes from 5×5×5 to 3×3×3

Experiments and Methodology

The paper explores different input configurations for segmentation, including:

• Types of input frames
• Inclusion of the annular curve
• Inclusion of commissural landmarks
• Resampling

All combinations of these modifications were tested across the different frame types, yielding 32 experiments in total (4 input frame types × 8 variations).

Input Frames

• Single-Phase: only the targeted mid-systolic (MS) frame
• Two-Phase: adds the mid-diastolic (MD) frame in addition to MS; provides information on valve opening
• Four-Phase: includes end-systolic (ES) and end-diastolic (ED) frames, resulting in 4 frames (ES, ED, MS, MD)
• Consecutive Systolic Phases (CSP): includes 10 frames around MS (5 before, 5 after), providing temporal context

Results

The best performance was achieved when using CSP frames with annular curve, commissural landmarks, and resampling — i.e., the maximum information available. This configuration outperformed human annotators.

• Individual leaflets: DSC = 0.81, MBD = 0.38 mm
• Merged valve: DSC = 0.85, MBD = 0.33 mm

Paper Breakdown

Category:
Clinical application research applying FCNs to segment tricuspid valve leaflets from 3D echocardiography in pediatric patients with congenital heart disease.

Context:
Addresses a critical need in pediatric cardiology where manual segmentation of tricuspid valves in HLHS patients is extremely time-consuming (2–4 hours) and highly variable between operators. While prior work focused on atlas-based approaches, this represents the first application of deep learning to pediatric congenital valve segmentation.

Correctness:
The methodology is sound with appropriate validation. However, the training dataset is imbalanced across surgical stages (58% post-stage 3 vs only 9% post-stage 2), which may affect generalizability.

Contributions:

• First deep learning approach for tricuspid valve segmentation in congenital heart disease
• Systematic evaluation of different input configurations (frame types, landmarks, preprocessing)
• Segmentation accuracy comparable to human intra-observer variability

Clarity:
The paper is well-structured, with clear methodology and comprehensive results. The clinical context is well-explained for a technical audience, though the 32 experimental configurations make the results section dense.

Additional Notes:

• Your observation about MBD being relevant for VR-Heart shell creation is insightful — boundary accuracy is indeed critical for haptic and simulation applications.
• The issue of signal dropout (where FCNs create holes that humans fill using anatomical knowledge) is a common challenge in medical AI and worth considering for cardiac applications.
• While the experimental design is strong, dataset limitations and class imbalance suggest cautious interpretation, particularly for underrepresented surgical stages.

Follow-up considerations:

• Use of MBD in VR-Heart applications
• Wilcoxon signed-rank test for pairwise statistical comparison
• Shapiro-Wilk test for normality