EU-Net: Automatic U-Net neural architecture search with differential evolutionary algorithm for medical image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Manual segmentation of medical images is time-consuming and error-prone.
• U-Net automates segmentation but requires expertise in neural network design.
• Optimizing U-Net architectures is crucial for accurate medical image analysis.

Purpose of the Study:

• To develop an automatic U-Net Neural Architecture Search (NAS) algorithm for medical image segmentation.
• To assist physicians in diagnosis by enhancing the accuracy of image interpretation.
• To automate the search for optimal U-Net architectures without requiring specialized expertise.

Main Methods:

• Proposed an automatic U-Net NAS algorithm named EU-Net, utilizing the differential evolutionary (DE) algorithm.
• Implemented a variable-length strategy for automatic architecture search.
• Incorporated DE's crossover, mutation, and selection strategies for exploration-exploitation balance.
• Introduced block-based and layer-based structures in encoding/decoding phases for optimization.

Main Results:

• EU-Net demonstrated superior performance on CHAOS and BUSI medical image segmentation datasets.
• The algorithm successfully automated U-Net architecture search, reducing the need for expert knowledge.
• Achieved at least a 6% improvement in the mean Intersection over Union (mIoU) metric compared to the original U-Net.

Conclusions:

• EU-Net effectively automates the optimization of U-Net architectures for medical image segmentation.
• The proposed method enhances diagnostic accuracy and efficiency in clinical practice.
• EU-Net offers a promising solution for complex medical image analysis tasks.