Publication: Detection of COVID-19 cases from chest x-ray images using pre-trained convolutional neural networks
Loading...
Date
2024-10
Authors
Ng, Tarng Wing
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The COVID-19 pandemic has highlighted the urgent need for efficient diagnostic tools. Traditional methods like Reverse Transcription-Polymerase Chain Reaction (RT-PCR) are often slow and resource-intensive. This study investigates deep learning approaches for early COVID-19 detection using chest X-ray images, proposing a hierarchical classification framework with convolutional neural networks (CNNs) to identify COVID-19, viral pneumonia, and normal cases. This study utilized a dataset of 18090 chest X-ray images split into training (14468), testing (1811), and validation (1811) sets. The hierarchical framework
consists of two stages: the first stage classifies normal and infected cases, while the second stage classifies infected cases into COVID-19 and viral pneumonia. Three pre-trained CNN models were evaluated: VGG-16, Inception V3, and ResNet-50. The VGG-16 model achieved test accuracy of 97.52% in the first stage and 99.87% in the second stage. The Inception V3 model performed best, achieving 99.01% accuracy in the first stage and 99.87% in the second stage. The ResNet-50 model achieved a test accuracy of 76.86% in the first stage and 92.30% in the second stage. Using transfer learning, the models were fine-tuned and optimized for the dataset. Standard evaluation metrics, including accuracy and confusion matrices, were used for comparison. The results indicate that the Inception V3 model is the most accurate, followed by VGG-16, which significantly outperforms ResNet-50. The Inception V3 model also demonstrated the best performance in training time and GPU RAM usage, making it the most efficient choice. Implementing these models in clinical settings can speed up the diagnostic
process, enabling quicker isolation and treatment of COVID-19 patients. This study demonstrates the effectiveness of deep learning models in medical image classification, offering a promising solution for enhancing COVID-19 detection accuracy and speed.