A lightweight network for COVID-19 detection in X-ray images

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Disease Detection
• Radiology and Diagnostic Imaging

Background:

• The Novel Coronavirus 2019 (COVID-19) pandemic necessitates rapid and accurate diagnostic tools.
• Current COVID-19 diagnosis relies on molecular tests and medical imaging like chest X-rays (CXR).
• Manual interpretation of CXR for COVID-19 is time-consuming and labor-intensive, especially during high demand.

Purpose of the Study:

• To develop and implement lightweight deep learning networks for efficient COVID-19 detection in CXR.
• To address the computational and memory demands of existing large neural networks for rapid diagnosis.
• To improve the speed and reduce the cost of COVID-19 diagnosis using automated CXR analysis.

Main Methods:

• Data augmentation of CXR images based on expert-defined clinical visual features.
• Design of a targeted, four-layer lightweight neural network architecture using 11x11 or 3x3 kernels.
• Implementation of a weight importance-based pruning criterion to further optimize network efficiency.

Main Results:

• The proposed lightweight network effectively recognizes regional and detail features in CXR for COVID-19 detection.
• Experimental validation on a public COVID-19 dataset demonstrated the method's effectiveness and efficiency.
• The optimized network achieved rapid yet accurate COVID-19 diagnosis, outperforming computationally intensive models.

Conclusions:

• Lightweight deep learning networks offer a promising solution for rapid and cost-effective COVID-19 diagnosis from CXR.
• The developed targeted network architecture and pruning strategy significantly reduce computational load.
• This approach can alleviate the strain on healthcare systems by enabling faster and more accessible diagnostic capabilities.