EMONAS-Net: Efficient multiobjective neural architecture search using surrogate-assisted evolutionary algorithm for...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning is vital for medical image segmentation but manual network design is complex and time-consuming.
• Existing multi-objective neural architecture search (NAS) methods have limitations in handling volumetric data and search efficiency.
• Over-parameterized and task-specific networks are common due to the challenges in automated design.

Purpose of the Study:

• To present EMONAS-Net, an efficient multi-objective NAS framework for 3D medical image segmentation.
• To optimize both segmentation accuracy and network size automatically.
• To address limitations of existing NAS methods in terms of search space and efficiency for volumetric data.

Main Methods:

• Developed EMONAS-Net with a novel search space considering micro- and macro-architecture structures.
• Employed a Surrogate-assisted Multiobjective Evolutionary based Algorithm (SaMEA) for efficient hyperparameter optimization.
• Integrated a Random Forest surrogate model to accelerate fitness evaluation of candidate architectures.

Main Results:

• EMONAS-Net achieved superior or comparable performance to state-of-the-art NAS methods on prostate, hippocampus, and cardiac segmentation tasks.
• The framework generated considerably smaller network architectures.
• Architecture search time was reduced by more than 50% across benchmarks.

Conclusions:

• EMONAS-Net provides an efficient and effective automated approach for designing 3D medical image segmentation networks.
• The framework successfully balances segmentation accuracy and network size.
• SaMEA algorithm and surrogate modeling significantly improve NAS efficiency for volumetric medical imaging.