DeepCOVIDNet-CXR: deep learning strategies for identifying COVID-19 on enhanced chest X-rays

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• COVID-19 remains a significant global health threat, necessitating accurate diagnostic tools.
• Deep learning models show promise for COVID-19 detection from chest X-rays, but limited data presents a challenge.
• Enhancing image quality is crucial for improving the performance of deep learning models in medical diagnostics.

Purpose of the Study:

• To evaluate COVID-19 identification performance using adaptive histogram equalization (AHE) with Convolutional Neural Network (ConvNet) architectures.
• To optimize AHE parameters for improved lung anatomy visualization in chest X-rays for COVID-19 detection.
• To assess the effectiveness of transfer learning on various ConvNet models for COVID-19 classification.

Main Methods:

• Experimentation with balanced small- and large-scale COVID-19 chest X-ray datasets.
• Application of adaptive histogram equalization (AHE) with varied parameters on lung and complete chest X-rays.
• Transfer learning applied to four ConvNet architectures: MobileNet, DarkNet19, VGG16, and AlexNet.

Main Results:

• DarkNet19 achieved 98.26% accuracy on small-scale datasets for multi-case identification.
• VGG16 demonstrated superior generalization, reaching 95.04% accuracy on large-scale datasets.
• Specific AHE parameters were identified as crucial for ConvNet performance.

Conclusions:

• This study pioneers the analysis of 3615 COVID-19 cases using AHE and ConvNets.
• The research specifies optimal AHE parameters for ConvNet-based multi-case classification of COVID-19.
• The findings contribute to more reliable AI-driven diagnostic tools for COVID-19 using chest radiography.