Analyze COVID-19 CT images based on evolutionary algorithm with dynamic searching space | JoVE Visualize

Area of Science:

Artificial Intelligence
Medical Imaging
Computer Vision

Background:

The COVID-19 pandemic necessitated rapid diagnostic tools, highlighting limitations in traditional testing and hospital resource allocation.
Convolutional Neural Networks (CNNs) have advanced medical image analysis, but their manual design requires scarce expertise.
Evolutionary Algorithms (EAs) offer automated CNN architecture search and hyperparameter optimization, reducing reliance on expert knowledge.

Purpose of the Study:

To propose a novel Evolutionary Algorithm (EA)-based method for automatically designing optimal Convolutional Neural Network (CNN) architectures for COVID-19 diagnosis from CT images.
To evaluate the performance of EA-designed CNNs against state-of-the-art models on the COVID-CT dataset.
To investigate the impact of specific architectural choices, such as batch normalization, on diagnostic performance.

Main Methods:

Development of a novel EA-based algorithm incorporating a dynamic search space for CNN architecture optimization.
Training and evaluation of the proposed EA-designed CNN on the COVID-CT dataset.
Comparative analysis against established CNN models, assessing performance without image preprocessing.

Main Results:

The EA-optimized CNN architecture achieved superior diagnostic performance compared to existing state-of-the-art models on the COVID-CT dataset.
The proposed method successfully diagnosed COVID-19 from CT images without requiring any preprocessing operations.
Experimentation indicated that excessive use of batch normalization might negatively impact CNN performance in this context.

Conclusions:

Automated CNN architecture design using EAs provides an effective approach for COVID-19 diagnosis from CT scans, bypassing the need for manual expertise and preprocessing.
The findings challenge conventional practices in CNN design, suggesting that intensive batch normalization may not always be beneficial.
This research offers a valuable tool for enhancing diagnostic capabilities and assisting experts in developing high-performing CNN models for medical imaging applications.