Related Experiment Video
Updated: Jun 9, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
370
Dual-attention transformer-based hybrid network for multi-modal medical image segmentation
Menghui Zhang1, Yuchen Zhang1, Shuaibing Liu1
1Department of Hepatobiliary and Pancreatic Surgery, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, 450001, China.
Scientific Reports
|October 28, 2024
Summary
A new deep learning model, DATTNet (Dual Attention Network), enhances medical image segmentation by effectively combining global and local feature extraction. This novel approach shows superior performance across various medical imaging tasks and modalities.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Accurate medical image segmentation is crucial for clinical diagnosis and treatment planning.
- Existing deep learning models like Convolutional Neural Networks (CNNs) and Transformers have limitations in capturing both global dependencies and local details.
- There is a need for advanced models that can effectively integrate multi-scale features for improved segmentation accuracy.
Purpose of the Study:
- To propose DATTNet, a Dual Attention Network, for robust medical image segmentation.
- To address the limitations of CNNs and Transformers in modeling global dependencies and local details, respectively.
- To evaluate the performance and generalizability of DATTNet across diverse medical imaging datasets and segmentation tasks.
Main Methods:
- Developed DATTNet, an encoder-decoder deep learning model incorporating a Dual Attention module for spatial and channel-wise global dependency modeling.
- Introduced a Context Fusion Bridge to remix multi-scale feature maps and establish feature correlations.
- Conducted experiments on ACDC, Synapse, and Kvasir-SEG datasets for segmentation of cardiac, abdominal organs, and gastrointestinal polyps.
Main Results:
- DATTNet achieved superior performance compared to state-of-the-art methods.
- Achieved high mean Dice Similarity Coefficient scores: 92.2% (cardiac), 84.5% (abdominal organs), and 89.1% (gastrointestinal polyps).
- Demonstrated favorable capability across different modalities (MRI, CT, endoscopy) and tasks, indicating strong generalizability.
Conclusions:
- DATTNet effectively models global dependencies and local details, leading to enhanced medical image segmentation.
- The model's robustness and generalizability suggest its potential for practical clinical applications.
- The proposed Dual Attention module and Context Fusion Bridge offer significant advancements in deep learning for medical imaging.
