Multi-scale conv-attention U-Net for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• U-Net networks are standard for medical image segmentation.
• Challenges exist in capturing multi-scale features and spatial context for complex structures.

Purpose of the Study:

• To propose a novel U-Net based network, MSCA-UNet, for enhanced medical image segmentation.
• To improve feature expression and segmentation performance by integrating specialized modules.

Main Methods:

• Developed MSCA-UNet integrating Adaptive Convolution (AC), Multi-Scale Learning (MSL), and Conv-Attention modules.
• AC module adaptively extracts features of varying shapes and scales.
• MSL module fuses multi-resolution features for rich connections.
• Conv-Attention module captures global context efficiently in skip connections.

Main Results:

• MSCA-UNet demonstrated significant improvements in segmentation accuracy and model robustness.
• Validated on CVC-ClinicDB, MICCAI 2023 Tooth, and ISIC2017 datasets.
• Outperformed existing segmentation methods while maintaining a lightweight architecture.

Conclusions:

• The proposed MSCA-UNet effectively addresses limitations in multi-scale feature and context extraction for medical image segmentation.
• Achieved superior performance and robustness, offering a lightweight and efficient solution.