Feature Separation in Diffuse Lung Disease Image Classification by Using Evolutionary Algorithm-Based NAS.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computational Biology

Background:

• Neural networks (NNs) show promise in lung disease diagnosis via image classification.
• NNs are often 'black boxes,' hindering trust and development in intelligent medicine.
• Lack of interpretability in NNs leads to skepticism and compromises model reliability.

Purpose of the Study:

• To introduce Evolutionary Neural Architecture Search (EvoNAS) for enhancing NN interpretability and accuracy in lung disease diagnosis.
• To develop an AI model that can effectively differentiate between critical and redundant features in medical images.
• To improve the reliability and trustworthiness of AI in medical applications.

Main Methods:

• Utilized an Evolutionary Algorithm to explore and optimize Convolutional Neural Networks (CNNs).
• Incorporated a Differential Evolution algorithm for enhanced search efficiency.
• Employed visualization techniques to ensure model interpretability.

Main Results:

• EvoNAS optimized CNNs excel at identifying discriminative features, improving classification accuracy.
• Achieved a 0.56% increase in classification accuracy on the diffuse lung disease dataset compared to the original network.
• Demonstrated significant advantages over existing methods for lung disease texture classification.

Conclusions:

• EvoNAS enhances the accuracy and interpretability of neural networks for lung disease diagnosis.
• The approach effectively distinguishes critical diagnostic features, improving classification performance.
• EvoNAS offers a more reliable and transparent AI solution for medical image analysis.