Combining external-latent attention for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Attention mechanisms enhance medical image segmentation.
• Current methods focus on external features, neglecting latent information.
• Existing attention schemes like global squeezing and self-attention have limitations.

Purpose of the Study:

• To propose TransGuider, an external-latent attention collaborative guided image segmentation network.
• To address the limitations of current attention mechanisms in medical image segmentation.
• To improve the exploitation of both external and latent features for better segmentation performance.

Main Methods:

• Developed a latent attention module using improved entropy quantification.
• Implemented an external self-attention module with sparse representation.
• Introduced a multi-attention collaborative module for refining segmentation masks.

Main Results:

• TransGuider demonstrated superior performance compared to state-of-the-art methods on benchmark datasets.
• Ablation experiments confirmed the effectiveness of individual components.
• The proposed method accurately explores latent contextual information and preserves external features.

Conclusions:

• TransGuider offers an effective approach for medical image segmentation by integrating external and latent attention.
• The network successfully leverages frequency changes within data for improved contextual understanding.
• The collaborative attention strategy refines segmentation masks, leading to enhanced accuracy.