Accurate cardiac ultrasound segmentation is essential for reliable assessment of ventricular function in intelligent healthcare systems. However, echocardiographic images are challenging due to low contrast, speckle noise, irregular boundaries, and domain shifts across devices and patient populations. Existing methods, largely based on appearance-driven learning, often fail to preserve boundary precision and structural consistency under these conditions. To address these issues, we propose a Contour-Guided Query Refinement Network (CGQR-Net) for boundary-aware cardiac ultrasound segmentation. The framework integrates multi-resolution feature representations with contour-derived structural priors. An HRNet backbone preserves high-resolution spatial details while capturing multi-scale context. A coarse segmentation is first generated, from which anatomical contours are extracted and encoded into learnable query embeddings. These contour-guided queries interact with fused feature maps via cross-attention, enabling structure-aware refinement that improves boundary delineation and reduces noise artifacts. A dual-head supervision strategy jointly optimizes segmentation and boundary prediction to enforce structural consistency. The proposed method is evaluated on the CAMUS dataset and further validated on the CardiacNet dataset to assess cross-dataset generalization. Experimental results demonstrate improved segmentation accuracy, enhanced boundary precision, and robust performance across varying imaging conditions. These results highlight the effectiveness of integrating contour-level structural information with feature-level representations for reliable cardiac ultrasound segmentation.
Figures & tables
Fig. 1: Sample echocardiographic images from the CAMUS (top row) and CardiacNet (bottom row) datasets, illustrating differences in image quality, noise levels, and anatomical variability. The CAMUS dataset contains relatively clean echocardiographic images with well-defined cardiac structures, while the CardiacNet dataset exhibits significantly higher variability, including severe speckle noise, low contrast, and irregular anatomical boundaries.
Fig. 2: Architecture of the proposed CGQR-Net. HRNet extracts multi-resolution features from the input echocardiography image, and a coarse segmentation head provides an initial structural prediction. Contours extracted from the coarse mask are converted into query embeddings and used to refine fused multi-scale features through cross-attention. The refined representation is then passed to segmentation and boundary heads to produce the final boundary-aware multi-class segmentation.
Paper
Methodology
Drawbacks
U-Net [ 11 ]
Encoder–decoder CNN with skip connections for pixel-wise segmentation.
Relies on local features and often produces smooth or inaccurate boundaries in low-contrast ultrasound images.
Attention U-Net [ 7 ]
Introduces attention gates to focus on relevant regions.
Improves region localization but still lacks explicit structural modeling; boundary precision remains limited.
UNet++ [ 8 ]
Dense skip connections for improved multi-scale feature fusion.
Increases architectural complexity while remaining primarily driven by pixel-wise learning without explicit contour priors.
nnU-Net [ 26 ]
Self-configuring segmentation framework with adaptive design.
Provides a strong baseline but remains appearance-driven and struggles with domain shift and boundary ambiguity.
Produces smoother predictions but lacks precise boundary delineation in challenging regions.
Extended nnU-Net [ 28 ]
Improved nnU-Net with enhanced configurations and training strategies.
Offers improved optimization but still shows limited structural awareness and sensitivity to anatomical variability.
TABLE I: Summary of existing cardiac segmentation methods, their methodologies, and limitations.
Fig. 3: Overall workflow of the proposed framework. CAMUS data, including two-chamber (2CH) and four-chamber (4CH) echocardiographic views, are used for training and internal validation, while CardiacNet is used for external validation. After preprocessing, images are fed into the boundary-aware segmentation model, which integrates HRNet, multi-scale feature fusion, and contour-guided query refinement. The model is evaluated using Dice similarity coefficients and qualitative analysis, producing final multi-class segmentation outputs.
Fig. 4: Overview of the proposed contour-guided query refinement framework. A two-dimensional echocardiography image is first processed by a HRNet backbone to extract multi-resolution features, and a coarse segmentation map is then generated. Contour points extracted from the coarse prediction are converted into query embeddings. These contour-derived queries interact with fused multi-scale features through cross-attention, where the queries provide structural guidance and the fused features provide spatial context. The refined representation is finally passed to dual prediction heads to produce the segmentation output and boundary map. Here, 2D denotes two-dimensional, Conv convolution, BN batch normalization, and ReLU rectified linear unit.
Parameter
Value
Framework
PyTorch
Optimizer
AdamW
Initial Learning Rate
1×10−4
Learning Rate Scheduler
Cosine Annealing
Batch Size
4
Number of Epochs
100
TABLE II: Training hyperparameters used for model training.
L=Lseg+λLboundary
Algorithm 1 Phase-Specific Training of CGQR-Net
Model
Endocardium
Epicardium
Left Atrium
Avg.
U-Net [ 11 ]
93.69
86.25
85.34
88.43
Attention U-Net [ 7 ]
93.72
86.54
86.02
88.76
UNet++ [ 8 ]
93.81
86.88
86.47
89.05
nnU-Net [ 26 ]
93.33
87.06
85.13
88.51
CANet [ 27 ]
93.91
87.52
87.01
89.48
Extended nnU-Net [ 28 ]
92.90
85.81
86.54
88.42
TABLE III: Comparison of cardiac structure segmentation performance on the CAMUS dataset in terms of DSC (%). The best results are shown in bold.
Model
Left Atrium
Right Atrium
Left Ventricle
Right Ventricle
Avg.
U-Net [ 11 ]
88.61
87.98
91.51
88.19
89.07
Attention U-Net [ 7 ]
88.94
88.31
91.63
88.54
89.36
UNet++ [ 8 ]
89.05
88.42
91.68
88.71
89.47
nnU-Net [ 26 ]
88.66
88.47
91.40
88.50
89.26
CANet [ 27 ]
88.95
88.91
91.70
89.26
89.71
Extended nnU-Net [ 28 ]
88.21
88.23
91.03
88.28
88.94
TABLE IV: Comparison of segmentation performance on the CardiacNet dataset in terms of DSC (%). The best results are shown in bold.
Fig. 5: Qualitative segmentation results on the CAMUS dataset. From left to right: original image, ground truth, coarse prediction, contour points, boundary map, final prediction, and overlay visualization. The proposed method produces more accurate and sharper boundaries compared to the coarse segmentation.
Fig. 6: Qualitative segmentation results on the CardiacNet dataset. From left to right: original image, ground truth, coarse prediction, contour points, boundary map, final prediction, and overlay visualization. The proposed method shows improved robustness under challenging conditions, producing more accurate and consistent segmentations with better boundary delineation.
Fig. 7: Visual comparison of segmentation results with existing methods on the CAMUS dataset. From left to right: original image and predictions from different methods, including the proposed CGQR-Net. The proposed method produces more accurate segmentation with improved boundary delineation and better shape consistency compared to previous approaches.
Method
LV Endo.
LV Epi.
LA
Full CGQR-Net
93.95
88.09
89.15
w/o Boundary Head
93.21
87.34
88.42
w/o Coarse Head
93.48
87.76
88.63
w/o Contour Queries
92.87
86.95
87.88
w/o Feature Pyramid Fusion
92.65
86.71
87.54
w/o Teacher Forcing
93.02
87.11
88.06
TABLE V: Class-wise ablation study of CGQR-Net on the CAMUS dataset (% DSC)
Accurate segmentation of cardiac structures in cardiovascular magnetic resonance (CMR) images is essential for reliable diagnosis and treatment of cardiovascular diseases. However, manual segmentation remains time-consuming and suffers from significant inter-observer variability. Recent advances in deep learning, particularly foundation models such as the Segment Anything Model (SAM), demonstrate strong generalization but often lack the boundary precision required for clinical applications. To address this limitation, we propose CardioSAM, a hybrid architecture that combines the generalized feature extraction capability of a frozen SAM encoder with a lightweight, trainable cardiac-specific decoder. The proposed decoder introduces two key innovations: a Cardiac-Specific Attention module that incorporates anatomical topological priors, and a Boundary Refinement Module designed to improve tissue interface delineation. Experimental evaluation on the ACDC benchmark demonstrates that CardioSAM achieves a Dice coefficient of 93.39%, IoU of 87.61%, pixel accuracy of 99.20%, and HD95 of 4.2 mm. The proposed method surpasses strong baselines such as nnU-Net by +3.89% Dice and exceeds reported inter-expert agreement levels (91.2%), indicating its potential for reliable and clinically applicable cardiac segmentation.
Graph-based cardiac segmentation with implicit anatomical correspondences provides topological guarantees and population-level analysis capabilities, but models trained on independent frames of image sequences exhibit temporal discontinuities that affect reliable clinical measurements, particularly in cardiac ultrasound. In this work, we introduce self-supervised temporal regularization as a post-training refinement stage that exploits the temporal coherence in image sequences to enforce consistent cardiac segmentation and motion estimation over time, without requiring per-frame annotations. By penalizing velocity and acceleration discontinuities across consecutive frames, our method achieves temporally consistent segmentations while maintaining the learned anatomical correspondences. We further leverage these correspondences to automatically map landmarks to the AHA 17-segment clinical standard, enabling standardized regional assessment and detection of pathological myocardial motion patterns. Validation on CAMUS dataset demonstrates the clinical utility of combining temporal consistency with automatic regional mapping. The code is publicly available at https://github.com/david-montalvoo/MaskHybridGNet-TempReg
David Montalvo-García, Nicolás Gaggion, María J. Ledesma-Carbayo +1
Biomedical Image Technologies, ETSI Telecomunicación, Universidad Politécnica de Madrid, Madrid, Spain · Centro de Investigación Biomédica en Red de Bioingeniería, Biomateriales y Nanomedicina (CIBER-BBN), Instituto de Salud Carlos III, Madrid, Spain · APOLO Biotech, Ciudad Autónoma de Buenos Aires, Argentina +2
Unsupervised domain adaptation approaches have recently succeeded in various medical image segmentation tasks. The reported works often tackle the domain shift problem by aligning the domain-invariant features and minimizing the domain-specific discrepancies. That strategy works well when the difference between a specific domain and between different domains is slight. However, the generalization ability of these models on diverse imaging modalities remains a significant challenge. This paper introduces UDA-VAE++, an unsupervised domain adaptation framework for cardiac segmentation. We show that the evidence lower bound of VAE-based domain adaptation penalizes the mutual information between the image and the latent space, including the structure information that segmentation relies on. Maximizing the structure mutual information removes this penalty on the structure information, while the remaining information is still compressed. To estimate this term, we develop a novel Structure Mutual Information Estimation (SMIE) block with a global estimator, a local estimator, and a prior information matching estimator to maximize the mutual information between the reconstruction and segmentation tasks. Specifically, we design a novel sequential reparameterization scheme that enables information flow and variance correction from the low-resolution latent space to the high-resolution latent space. Comprehensive experiments on benchmark cardiac segmentation datasets demonstrate that our model outperforms previous state-of-the-art qualitatively and quantitatively. The code is available at https://github.com/LOUEY233/Toward-Mutual-Information
Changjie Lu, Shen Zheng, Gaurav Gupta
College of Science and Technology, Wenzhou-Kean University, Wenzhou, China