Half-UNet: A Simplified U-Net Architecture for Medical Image Segmentation

Area of Science:

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

Background:

• Medical image segmentation is crucial for computer-aided diagnosis.
• U-Net is a popular architecture for medical image segmentation, with many variants.
• The fixed U-shaped structure of U-Net may not be optimal for all tasks.

Purpose of the Study:

• To experimentally analyze the impact of different U-Net components on segmentation performance.
• To propose a more efficient architecture, Half-U-Net, for medical image segmentation.
• To evaluate Half-U-Net against U-Net and its variants on diverse medical imaging datasets.

Main Methods:

• Experimental analysis of U-Net components.
• Development of the Half-U-Net architecture, a simplified encoder-decoder network.
• Incorporation of channel number unification, full-scale feature fusion, and Ghost modules.
• Comparative evaluation on mammography, lung CT, and cardiac MRI segmentation tasks.

Main Results:

• Half-U-Net demonstrates comparable segmentation accuracy to U-Net and its variants.
• Significant reduction in parameters (98.6%) and floating-point operations (81.8%) compared to U-Net.
• Validation across multiple medical image segmentation benchmarks.

Conclusions:

• The proposed Half-U-Net architecture provides an efficient alternative to U-Net for medical image segmentation.
• Half-U-Net achieves substantial computational savings without compromising segmentation accuracy.
• This simplified architecture holds promise for improving the efficiency of computer-aided diagnosis systems.