Multifaceted Disease Diagnosis: Leveraging Transfer Learning with Deep Convolutional Neural Networks on Chest X-Rays for COVID-19, Pneumonia, and Tuberculosis

Abstract

Introduction The three prevalent yet detrimental respiratory conditions, namely COVID-19, pneumonia, and tuberculosis, exhibit overlapping symptoms, making their differentiation challenging. However, their treatments are significantly divergent. Early detection emerges as a critical common factor for the effective management of these diseases. The pivotal initial step necessitates precise identification to initiate prompt prognosis. However, because of the lack of availability of experts in general and the inadequacy of the medical system on the whole, the problem of early detection is becoming highly concerning and, worst of all, time-consuming. Objective This research aimed to address this problem by examining and contrasting various deep Convolutional Neural Network (CNN) models that can accurately identify these illnesses, thereby assisting in their early detection. Methods 4 pre-trained CNN architectures have been used in this work, namely EfficientNet-B0, VGG-16, InceptionNet, and ResNet-50, which have been implemented on the input dataset. Firstly, the data were collected and pre-processed, and then model training and testing were performed for all 4 pre-trained models specified above. Results After fine-tuning the models and evaluating the test metrics on the test dataset, the highest accuracy was observed for ResNet-50 and EfficientNet models, i.e., ~95%. Also, the precision and recall for both were very similar (approximately greater than 92%), indicating accurate and good-quality results. Conclusion In this work, a transfer learning system has been employed utilizing several pre-trained CNN architectures. Our findings have indicated that this system can effectively analyze X-ray images to diagnose COVID-19, pneumonia, and tuberculosis.