ENAS-B: Combining ENAS With Bayesian Optimization for Automatic Design of Optimal CNN Architectures for Breast Lesion...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging Analysis

Background:

• Efficient Neural Architecture Search (ENAS) optimizes Convolutional Neural Network (CNN) cell structures but not entire architectures or hyperparameters.
• Existing ENAS methods have limitations in comprehensive CNN design for medical applications like breast lesion classification.

Purpose of the Study:

• To develop a novel framework for automatic CNN architecture design by integrating ENAS and Bayesian Optimization.
• To optimize both network architecture (cells, depth) and trainable hyperparameters for enhanced performance.

Main Methods:

• Utilized ENAS to identify optimal normal and reduction cells for CNNs.
• Employed Bayesian Optimization to determine optimal network depth and hyperparameter configurations.
• Validated the framework on a dataset of 1522 breast lesion ultrasound images and tested on external datasets.

Main Results:

• The proposed framework generated robust and computationally light CNN models.
• Achieved error rates of no more than 20.6% on internal tests and an average of 17.3% on external tests.
• Outperformed default ENAS-based CNNs and state-of-the-art architectures in classification tasks.

Conclusions:

• The combined ENAS and Bayesian Optimization framework effectively automates CNN architecture design.
• The approach yields high-performing, efficient models for breast lesion classification from ultrasound images.
• Demonstrates significant improvements over existing methods in robustness and performance.