Multi-Objective Once-for-All Neural Architecture Search for Medical Image Segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning dominates medical image segmentation.
• Existing neural architecture search (NAS) methods struggle with high computational costs and low-fidelity evaluations, limiting the discovery of efficient, high-performance networks.

Purpose of the Study:

• To develop a novel, efficient once-for-all NAS method for medical image segmentation.
• To address the limitations of current NAS approaches in finding lightweight and high-performance network architectures.

Main Methods:

• Designed a specialized search space (supernet) with U-shape network components for medical image segmentation.
• Implemented a hybrid two-stage supernet training scheme for balanced performance and computational cost.
• Utilized a multi-objective evolutionary algorithm to search for architectures optimized for multiple objectives.

Main Results:

• Achieved state-of-the-art performance on six diverse medical image segmentation datasets.
• The searched architectures demonstrated a superior trade-off between performance and computational complexity.
• Validated the effectiveness of the multi-objective search strategy.

Conclusions:

• The proposed once-for-all NAS method significantly advances medical image segmentation.
• This approach offers a practical solution for developing efficient and accurate deep learning models in medical imaging.
• The method provides a set of optimized architectures suitable for various computational constraints.