MSU-Net: Multi-Scale U-Net for 2D Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Traditional U-Net architectures face limitations with fixed receptive fields and undetermined optimal network widths.
• Extracting comprehensive semantic features is crucial for accurate medical image segmentation.

Purpose of the Study:

• To address U-Net's limitations by proposing a Multi-scale U-Net (MSU-Net).
• To enhance feature extraction diversity and alleviate issues with optimal network width in medical image segmentation.

Main Methods:

• Employed multiple convolution sequences for richer semantic feature extraction.
• Integrated convolution kernels with varying receptive fields to diversify features.
• Validated the multi-scale block's universality by extending it to other U-Net variants.

Main Results:

• MSU-Net achieved high Intersection over Union (IoU) scores across five diverse medical image segmentation datasets (0.771, 0.867, 0.708, 0.900, 0.702).
• Demonstrated superior performance compared to existing methods on datasets including electron microscopy, dermoscopy, and ultrasound images.
• The multi-scale block proved effective and universally applicable across different U-Net architectures.

Conclusions:

• MSU-Net effectively overcomes the limitations of fixed receptive fields in U-Net for medical image segmentation.
• The proposed architecture achieves state-of-the-art performance on various medical imaging datasets and modalities.
• The MSU-Net framework offers a robust and adaptable solution for diverse medical image segmentation tasks.