MFR-UNet: A Medical Image Segmentation Network With Fused Multi-Scale Feature Refinement

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Medical image segmentation is vital for clinical diagnosis and treatment planning.
• Current Convolutional Neural Network (CNN) based methods, like U-Net, face challenges in capturing long-range dependencies and integrating multi-level features.
• Existing models struggle with refining feature representations and efficiently fusing information across different network levels.

Purpose of the Study:

• To propose a novel U-Net architecture, the multi-scale feature refinement U-Net (MFR-UNet), to enhance medical image segmentation.
• To address limitations in capturing long-range dependencies, refining multi-level features, and integrating cross-level information.
• To improve segmentation accuracy, robustness, and boundary clarity in medical images.

Main Methods:

• Developed a novel U-Net architecture (MFR-UNet) incorporating three key modules: Wavelet Transform Convolution (WtConv), Large Receptive Field Attention (LRFA), and Weighted Contextual Fusion (WCF).
• WtConv module processes features in the frequency domain for precise learning of high-frequency details and low-frequency contours.
• LRFA module in the encoder uses deep separable convolutions and multi-head attention to efficiently capture global context.
• WCF module in skip connections and decoding path adaptively fuses feature streams using dynamic channel attention weights.

Main Results:

• The MFR-UNet demonstrated superior performance compared to several mainstream methods on multiple public medical image segmentation datasets.
• Achieved significant improvements in key segmentation metrics, including Dice coefficient and Intersection over Union (IoU).
• The proposed modules effectively enhanced segmentation accuracy and improved the clarity of segmented boundaries.

Conclusions:

• The MFR-UNet effectively addresses the limitations of existing U-Net variants in medical image segmentation.
• The integration of WtConv, LRFA, and WCF modules leads to enhanced feature representation and fusion, boosting segmentation performance.
• MFR-UNet shows significant potential for improving clinical diagnosis and treatment planning through accurate and robust medical image segmentation.