A novel recursive transformer-based U-Net architecture for enhanced multi-scale medical image segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Deep Learning Architectures

Background:

• Automatic medical image segmentation is crucial for diagnosis and treatment planning.
• U-Net is widely used but struggles with long-range dependencies in complex medical images.
• Anatomical structures can be difficult to distinguish from backgrounds in textured images.

Purpose of the Study:

• To introduce a novel network, the recursive transformer-based U-Net (ReT-UNet), to overcome U-Net's limitations.
• To enhance the capture of long-range dependencies and improve feature abstraction in medical images.
• To improve the accuracy and stability of medical image segmentation.

Main Methods:

• Developed ReT-UNet integrating recursive feature learning and transformer technology.
• Introduced a multi-scale global feature fusion (Multi-GF) module for enhanced contextual understanding.
• Incorporated a recursive feature accumulation block and a lightweight atrous spatial pyramid pooling (ASPP) module.

Main Results:

• Ablation experiments demonstrated consistent performance across multiple trials.
• The method showed reduced mis-segmented regions in cardiac, pulmonary nodule, and polyp segmentation.
• Improved segmentation accuracy and stability compared to existing state-of-the-art methods.

Conclusions:

• ReT-UNet demonstrates superior performance over related methods in medical image segmentation.
• The proposed architecture holds significant potential for various medical image analysis applications.