A global-frequency-domain network for medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Traditional UNet networks excel in medical image segmentation but suffer from high parameter counts and computational demands.
• Limited medical datasets pose challenges for training complex models like UNet.
• Need for efficient and effective medical image segmentation architectures.

Purpose of the Study:

• Introduce Global Frequency UNet (GFUNet), a novel architecture for fast medical image segmentation.
• Address the limitations of traditional UNet models regarding parameter count and computational complexity.
• Improve segmentation performance on limited medical datasets.

Main Methods:

• Developed GFUNet by combining Fourier Transform with UNet structure for efficient encoding and decoding.
• Incorporated a dual-domain encoding module to leverage frequency domain features.
• Inspired by modified Multi-Layer Perceptron (MLP)-like models for architectural innovation.

Main Results:

• GFUNet significantly reduces parameters (46x) and computational complexity (114x) compared to the original UNet.
• Achieved improved segmentation performance across various medical imaging tasks.
• Demonstrated enhanced efficiency and effectiveness in encoding and decoding processes.

Conclusions:

• GFUNet offers a highly efficient alternative for medical image segmentation, especially with limited data.
• The integration of Fourier Transform and dual-domain encoding enhances model performance and reduces resource requirements.
• GFUNet represents a significant advancement in fast and accurate medical image segmentation.