Detection of COVID-19, lung opacity, and viral pneumonia via X-ray using machine learning and deep learning.
Authors
Affiliations (5)
Affiliations (5)
- Laboratory of Condensed Matter and Interdisciplinary Sciences, Physics Department, Faculty of Sciences, Mohammed V University in Rabat, Morocco.
- Laboratory of Condensed Matter and Interdisciplinary Sciences, Physics Department, Faculty of Sciences, Mohammed V University in Rabat, Morocco; College of computing, Mohammed VI Polytechnic University, Lot 660, Hay Moulay Rachid Ben Guerir, 43150, Morocco.
- ENSAM-R - Université Mohammed V de Rabat, Morocco.
- Laboratory of Condensed Matter and Interdisciplinary Sciences, Physics Department, Faculty of Sciences, Mohammed V University in Rabat, Morocco; Hassan II Academy of Science and Technology, Rabat, Morocco.
- Department of Electrical and Communication Engineering, College of Engineering, UAE University, P.O. Box: 15551, Al Ain, United Arab Emirates. Electronic address: [email protected].
Abstract
The COVID-19 pandemic has significantly strained healthcare systems, highlighting the need for early diagnosis to isolate positive cases and prevent the spread. This study combines machine learning, deep learning, and transfer learning techniques to automatically diagnose COVID-19 and other pulmonary conditions from radiographic images. First, we used Convolutional Neural Networks (CNNs) and a Support Vector Machine (SVM) classifier on a dataset of 21,165 chest X-ray images. Our model achieved an accuracy of 86.18 %. This approach aids medical experts in rapidly and accurateky detecting lung diseases. Next, we applied transfer learning using ResNet18 combined with SVM on a dataset comprising normal, COVID-19, lung opacity, and viral pneumonia images. This model outperformed traditional methods, with classification rates of 98 % with Stochastic Gradient Descent (SGD), 97 % with Adam, 96 % with RMSProp, and 94 % with Adagrad optimizers. Additionally, we incorporated two additional transfer learning models, EfficientNet-CNN and Xception-CNN, which achieved classification accuracies of 99.20 % and 98.80 %, respectively. However, we observed limitations in dataset diversity and representativeness, which may affect model generalization. Future work will focus on implementing advanced data augmentation techniques and collaborations with medical experts to enhance model performance.This research demonstrates the potential of cutting-edge deep learning techniques to improve diagnostic accuracy and efficiency in medical imaging applications.