Contour-Guided Query-Based Feature Fusion for Boundary-Aware and Generalizable Cardiac Ultrasound Segmentation

1. Problem Statement

Cardiac ultrasound (echocardiography) segmentation is critical for assessing ventricular function but faces significant challenges due to the intrinsic nature of ultrasound imaging:

• Image Quality Issues: Low contrast, severe speckle noise, and weak/ambiguous anatomical boundaries.
• Anatomical Variability: Complex non-linear deformations of cardiac structures across the cardiac cycle and significant variability between patients.
• Domain Shift: Performance degradation when models trained on one dataset (e.g., specific devices or protocols) are applied to unseen datasets with different acquisition conditions.
• Limitations of Existing Methods:
  • CNNs (e.g., U-Net, HRNet): Rely heavily on local pixel-wise appearance, often producing overly smooth or irregular boundaries and lacking explicit anatomical consistency.
  • Transformers: Capture long-range dependencies but often lack geometric grounding, leading to boundary leakage or jagged edges.
  • Boundary-Aware Methods: Typically treat boundary prediction as an auxiliary task rather than a core driver for feature refinement, failing to fully utilize structural priors.

2. Methodology: CGQR-Net

The authors propose CGQR-Net (Contour-Guided Query Refinement Network), a framework that explicitly integrates structural priors into the feature refinement process. The architecture consists of three main stages:

A. Multi-Resolution Feature Extraction (HRNet Backbone)

Instead of standard encoder-decoder backbones that aggressively downsample, the model uses HRNet.

• Mechanism: Maintains parallel high-resolution feature streams throughout the network.
• Benefit: Preserves fine spatial details and boundary cues essential for contour extraction while fusing multi-scale semantic context.

B. Coarse Segmentation and Contour Extraction

1. Coarse Prediction: A segmentation head generates an initial "coarse" mask ($S_c$) from the deepest feature map.
2. Contour Extraction: Anatomical contours are extracted from the boundaries of the coarse mask.
3. Geometric Encoding: Contours are converted into compact query embeddings. These embeddings include geometric descriptors such as:
   • Centroid coordinates ($\mu_x, \mu_y$)
   • Area ($A_k$) and Perimeter ($P_k$)
   • Dispersion terms ($\sigma_x, \sigma_y$)
   • These are projected into a latent space to form structural tokens ($q_k$).

C. Contour-Guided Cross-Attention Refinement

This is the core innovation. The model treats the contour-derived embeddings as queries that interact with the fused multi-scale image features (keys and values) via Cross-Attention.

• Process:
  • Queries ($Q$): Represent structural priors (shape, location, extent).
  • Keys/Values ($K, V$): Represent the fused image features (texture, context).
  • Attention Mechanism: The network computes attention weights to determine how strongly each contour query attends to specific spatial locations in the feature map.
  • Refinement: The output is a modulation map that selectively enhances boundary-relevant regions and suppresses noise-induced artifacts, resulting in a refined feature representation ($F_{ref}$).

D. Dual-Head Supervision

The refined features are passed to two parallel heads:

1. Segmentation Head: Predicts the final class mask.
2. Boundary Head: Predicts the edge map.

• Loss Function: A joint loss function combines Multi-class Dice Loss (for region overlap) and Binary Cross-Entropy Loss (for boundary precision), enforced by a weighting parameter $\lambda$.

3. Key Contributions

1. Contour-Guided Query Mechanism: A novel approach that transforms segmentation contours into learnable query embeddings, providing explicit structural guidance during feature refinement.
2. Structure-Aware Feature Fusion: Integration of multi-resolution features with cross-attention, allowing the model to fuse local details with global context guided by anatomical geometry.
3. Boundary-Aware Dual-Head Strategy: A joint optimization strategy that enforces anatomical consistency by simultaneously optimizing region segmentation and boundary prediction.
4. Robust Generalization: Demonstrated strong cross-dataset generalization by leveraging structural invariance (contours) over appearance-based features, reducing sensitivity to domain shifts.

4. Experimental Results

The method was evaluated on two datasets: CAMUS (standard benchmark) and CardiacNet (external validation with pathological cases and higher variability).

• Quantitative Performance (CAMUS):
  • Achieved a mean Dice Similarity Coefficient (DSC) of 90.40%, outperforming state-of-the-art baselines (U-Net, Attention U-Net, nnU-Net, DAM-Seg).
  • Notable improvements in difficult regions:
    • LV Epicardium: 88.09% (vs. 88.00% for DAM-Seg).
    • Left Atrium: 89.15% (vs. 87.01% for CANet).
• Cross-Dataset Generalization (CardiacNet):
  • Achieved a mean DSC of 90.27% on the unseen CardiacNet dataset.
  • Demonstrated superior robustness against domain shift compared to CNNs and standard Transformers, maintaining high accuracy even in cases with severe speckle noise and irregular anatomies (e.g., Atrial Septal Defect).
• Qualitative Analysis:
  • Visual results show significantly sharper boundaries and reduced "leakage" artifacts compared to baselines.
  • The model successfully corrected coarse segmentation errors, particularly in low-contrast regions.
• Ablation Studies:
  • Removing the Contour-Guided Query mechanism caused the largest performance drop (e.g., LV Epicardium DSC dropped from 88.09% to 86.95%), confirming its critical role.
  • Removing boundary supervision or feature fusion also resulted in consistent performance degradation.

5. Significance and Impact

• Clinical Relevance: The method addresses the critical need for precise boundary delineation in cardiac ultrasound, which is essential for calculating clinical metrics like ejection fraction and wall thickness.
• Paradigm Shift: Moves beyond purely appearance-based learning to structure-guided learning. By using contours as queries, the model learns to prioritize anatomical consistency over noisy texture patterns.
• Real-World Applicability: The strong cross-dataset generalization suggests the model is suitable for deployment in diverse clinical environments with varying equipment and patient populations, a major hurdle for current AI solutions in medical imaging.
• Future Direction: Establishes a new framework for integrating geometric priors into Transformer-based and attention-driven segmentation architectures.