SelfReg-UNet: Self-Regularized UNet for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning Architectures
• Computer Vision

Background:

• UNet is a leading architecture for medical image segmentation.
• Existing UNet improvements often overlook in-depth analysis of learned patterns.
• Factors like asymmetric supervision and feature redundancy can impact UNet performance.

Purpose of the Study:

• To analyze UNet's learned patterns in medical image segmentation.
• To identify and address factors limiting UNet performance.
• To propose an improved UNet methodology for enhanced segmentation accuracy.

Main Methods:

• Investigated UNet's internal feature learning patterns.
• Proposed a method to balance supervision between UNet's encoder and decoder.
• Implemented feature distillation to reduce redundancy in feature maps.
• Integrated proposed enhancements as a plug-and-play module.

Main Results:

• The proposed method consistently improved standard UNet performance across four medical image segmentation datasets.
• Addressed issues of irrelevant feature learning due to asymmetric supervision.
• Reduced feature redundancy within the UNet architecture.
• Achieved performance gains with negligible computational overhead.

Conclusions:

• Balancing encoder-decoder supervision and reducing feature redundancy are key to improving UNet.
• The proposed feature distillation and supervision balancing method offers a simple yet effective enhancement for UNet.
• This approach provides a practical solution for boosting medical image segmentation accuracy using UNet.