Hierarchical agent transformer network for COVID-19 infection segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Accurate segmentation of COVID-19 infection is vital for patient management.
• Convolutional Neural Networks (CNNs) struggle with irregular lesion shapes.
• Transformer models offer global context but are computationally intensive and have suboptimal feature integration.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for COVID-19 lesion segmentation.
• To address the limitations of existing CNN and Transformer models in medical image segmentation.
• To propose the Hierarchical Agent Transformer Network (HATNet) for improved COVID-19 detection.

Main Methods:

• Proposed Hierarchical Agent Transformer Network (HATNet) with an encoder-bridge-decoder architecture.
• Utilized agent Transformer blocks with linear complexity and a diversity restoration module (DRM) for subtle feature detection.
• Integrated an improved pyramid pooling module (IPPM) for global context and a full-scale bidirectional feature pyramid network (FsBiFPN) with a border-refinement module (BRM) for edge precision.

Main Results:

• HATNet achieved Dice scores of 84.14% on COVID-19-CT-Seg and 81.22% on CC-CCII datasets.
• Demonstrated superior segmentation performance compared to state-of-the-art models.
• Showcased significant advantages in model parameters and computational complexity.

Conclusions:

• HATNet effectively balances segmentation accuracy and computational efficiency for COVID-19 lesion segmentation.
• The proposed model exhibits strong potential for clinical deployment in medical imaging.
• HATNet offers a promising advancement in automated analysis of CT scans for COVID-19.