MHSU-Net: A more versatile neural network for medical image segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Medical image segmentation is crucial in clinical settings.
• Convolutional Neural Network (CNN)-based methods, like U-Net, are widely used.
• Standard U-Net has limitations in feature extraction and can lose information during pooling.

Purpose of the Study:

• To enhance the feature extraction capabilities of U-Net for medical image segmentation.
• To minimize feature loss during the segmentation process.
• To improve the versatility of segmentation models for various clinical tasks.

Main Methods:

• Introduced a novel Multiscale Convolutional Block (MCB) for richer feature capture.
• Developed a Hybrid Down-sampling Block (HDSB) to replace max-pooling and reduce feature loss.
• Proposed a Context Information Fusion (CIF) module using atrous convolution and SKNet for context extraction.
• Integrated CIF with U-Net's Skip Connection to create the Skip Connection+ structure, forming the MHSU-Net.

Main Results:

• The proposed MHSU-Net demonstrated superior performance compared to U-Net and other state-of-the-art models.
• Evaluations on lung, cell contour, and pancreas datasets showed significant improvements across various metrics.
• MHSU-Net exhibits strong potential for practical clinical applications.

Conclusions:

• The novel modules effectively boost feature extraction and reduce segmentation-related feature loss.
• MHSU-Net offers enhanced performance and versatility for diverse medical image segmentation tasks.
• The proposed architecture shows promise for advancing clinical image analysis.