GrMoNAS: A granularity-based multi-objective NAS framework for efficient medical diagnosis.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Neural Architecture Search (NAS) automates medical image diagnostics but demands extensive computational resources and time.
• Existing NAS methods often struggle with efficiency and resource constraints, particularly in clinical settings.

Purpose of the Study:

• To introduce GrMoNAS, a novel framework for efficient and accurate neural architecture search in medical image diagnostics.
• To address the computational limitations of traditional NAS methods for medical applications.
• To balance diagnostic accuracy and computational efficiency using proxy datasets and multi-objective optimization.

Main Methods:

• GrMoNAS employs a two-phase approach: coarse granularity (reduced proxy dataset) and fine granularity (comprehensive validation).
• Utilizes proxy datasets for granularity transformation to accelerate architecture evaluation.
• Incorporates multi-objective optimization and Pareto frontier sorting to enhance accuracy and efficiency simultaneously.

Main Results:

• GrMoNAS achieved comparable or superior diagnostic precision across various medical scenarios (COVID-19, skin cancer, etc.) compared to traditional models and recent NAS approaches.
• Demonstrated significant enhancement in diagnostic efficiency, making it suitable for resource-limited hospitals.
• Effectively avoided local optima, showcasing robust performance in precision medical diagnosis.

Conclusions:

• GrMoNAS offers a computationally efficient and effective solution for automated medical image diagnostics.
• The framework's ability to balance accuracy and efficiency makes it highly valuable for clinical implementation.
• GrMoNAS shows significant potential for advancing precision medical diagnosis, especially in resource-constrained environments.