DATA AUGMENTATION AND TRANSFER LEARNING IN DETECTION OF COVID-19

Abstract

Global healthcare systems are facing unprecedented challenges as a result of the COVID-19 pandemic, which calls for quick and precise diagnostic instruments to stop its spread. In this study, we investigate the application of transfer learning in convolutional neural networks (CNNs) and the influence of data augmentation in de- tecting COVID-19 from chest X-ray images. We leverage a dataset comprising images from patients diagnosed with COVID-19 and images corresponding to non-COVID cases, and as part of a transfer learning approach, we utilize four well known CNN models: ResNet50, MobileNet V2, EfficientNet V2, and MobileNet V3. A key focus of our research is the systematic investigation of data augmentation factors and their impact on model performance. Through varying the intensity and types of data aug- mentations, such as rotations, flipping and zooming, we seek to optimize the models’ ability to generalize from training data to real-world scenarios. Our findings reveal that precise calibration of data augmentation significantly en hances the diagnostic capabilities of the models. While increased augmentation gen erally improves sensitivity and specificity, excessive augmentation diminishes mod- els’ performances due to overfitting on non-realistic features. As another result, MobileNet V2 and MobileNet V3 show the highest specificity scores of 0.72 and 0.70, while EfficientNet V2 demonstrates superior sensitivity of 0.96, indicating the strengths and trade-offs of different architectures. These results, assessed through a comprehensive set of metrics including accuracy, sensitivity, specificity, precision, recall, F1 score, and AUC-ROC, underscore the effectiveness of deep learning meth- ods in COVID-19 identification and the crucial role of tailored data augmentation in improving model robustness. The implications of our results extend to clinical prac- tice and public health, highlighting the potential of integrating advanced machine learning technologies into healthcare workflows to enhance diagnostic efficiency and patient care. Looking ahead, we propose further exploration into additional imaging modalities, the integration of multi-modal data, and more sophisticated data aug- mentation techniques, such as usage of Generative Adversarial Networks (GANs), to refine diagnostic accuracy. Overall, our study reinforces the significance of transfer learning and deep learning in addressing the urgent challenges posed by infectious diseases like COVID-19, paving the way for more sophisticated diagnostic tools.