TA-Net: Triple attention network for medical image segmentation

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computer-aided diagnosis

Background:

• Convolutional Neural Networks (CNNs) and attention mechanisms have advanced medical image segmentation.
• Existing methods often focus on single-dimension attention, potentially missing cross-dimensional feature correlations.
• Capturing comprehensive global features across multiple dimensions remains a challenge.

Purpose of the Study:

• To propose a novel Triple Attention Network (TA-Net) for medical image segmentation.
• To address the limitation of single-dimension feature attention by integrating channel, spatial, and feature internal attention.
• To enhance the model's ability to capture global contextual information.

Main Methods:

• Developed a Triple Attention Network (TA-Net) incorporating channel, spatial, and feature internal attention domains.
• Introduced a Channel with Self-Attention Encoder (CSE) block to learn long-range pixel dependencies and increase receptive fields.
• Proposed a Spatial Attention Up-sampling (SU) block for improved fusion of low-level and high-level features during decoding.

Main Results:

• TA-Net demonstrated superior performance compared to state-of-the-art methods across diverse medical image segmentation tasks.
• Achieved high accuracy, robustness, and reduced redundancy in segmenting retinal and intracranial blood vessels, cells, and cutaneous melanoma.
• Experiments conducted on five datasets (DRIVE, STARE, ISBI 2012, ISIC 2017, and a local dataset) validated the model's effectiveness.

Conclusions:

• The proposed TA-Net effectively captures global contextual information by leveraging multi-dimensional attention mechanisms.
• TA-Net offers a promising approach for accurate and robust medical image segmentation.
• The integration of CSE and SU blocks enhances feature representation and spatial localization for improved segmentation outcomes.