HEA-Net: Attention and MLP Hybrid Encoder Architecture for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Transformer models excel at capturing global dependencies via self-attention but struggle with local details crucial for medical image segmentation.
• Accurate identification of blurred boundaries in medical images requires robust modeling of local features, a limitation in standard Transformer architectures.

Purpose of the Study:

• To propose HEA-Net, an attention and MLP hybrid-encoder architecture designed to overcome Transformer limitations in medical image analysis.
• To enhance the capture of local details and foreground information while suppressing background noise in medical images.
• To improve the accuracy of medical image segmentation, particularly in cases with blurred boundaries.

Main Methods:

• Developed HEA-Net, integrating an Efficient Attention Module (EAM) with a Dual-channel Shift MLP (DS-MLP) module.
• EAM connects convolutional blocks with Transformer to enhance foreground information and reduce background noise.
• DS-MLP further refines foreground details through channel and spatial shift operations.

Main Results:

• HEA-Net demonstrated superior performance on public medical image datasets.
• Achieved a Dice score of 90.56% and IoU of 83.62% on the GlaS dataset.
• Achieved a Dice score of 80.80% and IoU of 68.26% on the MoNuSeg dataset.

Conclusions:

• HEA-Net effectively combines global and local feature extraction for improved medical image segmentation.
• The proposed architecture successfully addresses the limitations of standard Transformers in handling local details and blurred boundaries.
• Experimental results validate the excellent performance and potential of HEA-Net for medical image analysis tasks.