What is the clinical evidence from this study?

Study design: Other. Population: Echocardiography image segmentation (n=500). Intervention: Multi-attention enhanced encoder-decoder network with hybrid transformer bottleneck vs. UNet, UNETR-2d, and Swin-unet. Primary outcome: Dice Similarity Coefficient (DSC) for left ventricle endocardium (LVendo) in four-chamber (4CH) view.

What does this research mean for the field?

A novel encoder-decoder network combining a Vision Transformer and a Multi Receptive Field Block with multiple attention mechanisms achieves superior segmentation accuracy for 2D echocardiograms compared to pure CNN, pure transformer, and existing hybrid architectures. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

Develop a novel deep learning architecture that enhances segmentation accuracy in 2D echocardiograms by combining CNNs and transformers.

July 10, 2026Open Access

Multi-attention enhanced encoder–decoder network with hybrid transformer bottleneck for echocardiography image segmentation

SCSaeed ChamaniIran University of Science and Technology HBHamid BehnamIran University of Science and Technology

Key Result

The proposed multi-attention enhanced encoder-decoder network achieved superior segmentation accuracy, reaching a Dice Similarity Coefficient of 92.25% for the left ventricle endocardium in four-chamber views, outperforming pure CNN and transformer-based models.

Key Points

Develop a novel deep learning architecture that enhances segmentation accuracy in 2D echocardiograms by combining CNNs and transformers.
Proposed an encoder-decoder framework with multiple attention mechanisms and hybrid bottleneck combining Vision Transformer (ViT) and Multi Receptive Field Block (MRFB)
Utilized Atrous Spatial Pyramid Pooling (ASPP) for deep supervision
Trained and validated on the CAMUS dataset.
Achieved Dice Similarity Coefficients of 91.11±0.19 for left ventricle endocardium, 87.60±0.11 for epicardium, and 87.85±0.29 for left atrium in two-chamber views
Achieved Dice Similarity Coefficients of 92.25±0.17, 87.29±0.17, and 91.71±0.10 in four-chamber views
Outperformed UNet, Swin-unet, and UNETR-2d architectures in segmentation accuracy.

Structured PICO

Population

500 clinically acquired 2D echocardiography images from the CAMUS dataset used to train and evaluate a deep learning segmentation model.

Intervention

Novel encoder-decoder framework with multiple attention mechanisms and hybrid bottleneck combining Vision Transformer (ViT) and Multi Receptive Field Block (MRFB)

Comparator

Pure CNN-based (UNet), pure transformer-based (Swin-unet), and CNN-Transformer hybrid (UNETR-2d) architectures

Outcome

Segmentation accuracy (Dice Similarity Coefficient) for left ventricle endocardium, epicardium, and left atrium in two-chamber and four-chamber viewssurrogate

A novel hybrid CNN-transformer deep learning architecture improves the accuracy of automatic 2D echocardiogram segmentation compared to existing models.

Main Result

Absolute Event Rate: 92.25% vs 91.81%

Limitations

Accurate segmentation remains difficult in the presence of significant image artifacts and poor contrast (worst-case scenarios).

Abstract

Automatic segmentation of 2D echocardiograms is a critical step in diagnosing and assessing heart diseases. Convolutional Neural Networks (CNNs), such as U-Net, have proven effective for medical image segmentation, but they often fall short in capturing long-range dependencies because convolution operations are inherently local. While transformer models capture the global context, they also may not be good in fine-grained localization. This study aims to develop a novel deep learning architecture that addresses these limitations by combining the strengths of CNNs and transformers for improved segmentation accuracy in 2D echocardiograms. We propose a novel encoder-decoder framework enriched with multiple attention mechanisms and hybrid bottleneck combining a Vision Transformer (ViT) and a Multi Receptive Field Block (MRFB). The integration of a ViT enables the modeling of long-range dependencies, while the MRFB improves multi-scale feature extraction and is specifically designed to improve the representation of spatial features and reduce the loss of spatial information. Enhanced by multiple attention mechanisms and deep supervision with Atrous Spatial Pyramid Pooling (ASPP), our model robustly captures both local and global contexts. Trained and validated on the CAMUS dataset, it achieves Dice Similarity Coefficients of 91. 11±0. 19, 87. 60±0. 11, and 87. 85±0. 29 for left ventricle endocardium, epicardium, and left atrium in two-chamber views, and 92. 25±0. 17, 87. 29±0. 17, and 91. 71±0. 10 in four-chamber views. Our model outperforms pure CNN-based (such as UNet), pure transformer-based (such as Swin-unet), and CNN-Transformer hybrid (such as UNETR-2d) architectures, achieving superior segmentation accuracy and robustness. Code is publicly available on the paper's https: //github. com/saeedchamani/CAMUSSegmentationHybridBottleneck repository.

Ask AI

Helpful

Bookmark

View Full Paper