Transformer and group parallel axial attention co-encoder for medical image segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• U-Net is a standard for medical image segmentation but struggles with long-range dependencies.
• Transformers excel at long-range dependencies but lack focus on local foreground information.
• Medical images require both local and global contextual understanding for accurate segmentation.

Purpose of the Study:

• To propose GPA-TUNet, a hybrid model addressing U-Net's limitations in long-term dependencies and Transformer's lack of local focus.
• To enhance medical image segmentation by effectively combining local and global information processing.

Main Methods:

• Introduced Group Parallel Axial Attention (GPA) to emphasize local foreground information.
• Integrated GPA with Transformer in the encoder for improved foreground highlighting and background noise reduction.
• Incorporated sMLP blocks to enhance the network's global modeling capabilities through sparse connectivity and weight sharing.

Main Results:

• GPA-TUNet demonstrated superior performance on public medical image segmentation datasets.
• Achieved a mean Dice Similarity Coefficient (DSC) of 80.37% on the Synapse dataset.
• Achieved a mean DSC of 90.37% and a mean Hausdorff Distance 95 (HD95) of 1.23 mm on the ACDC dataset.

Conclusions:

• GPA-TUNet effectively integrates local and global information for enhanced medical image segmentation.
• The proposed GPA mechanism and sMLP blocks contribute to improved accuracy and robustness.
• The model shows significant potential for clinical applications in medical image analysis.