Comparison Study Of Deep-Learning Architectures For Classification of Thoracic Pathology

Abstract

This work aims to study different architectures for the classification of thoracic diseases using pre-trained convolutional neural networks (PCNN) such as VGG-16, ResNet-50, EfficientNetB0, and InceptionV3 which are considered as state-of-the-art deep learning models. Indeed, they are used to detect various thoracic disorders. In this study, the main focus is on COVID-19 and pneumonia to make an optimal diagnosis for these two diseases. Although these diseases are prevalent, the process of detection and diagnosis is challenging. In this work, two unbalanced datasets (COVID-19 and Pneumonia) have been used. After the training phase where hyperparameters of the models have been tuned for best accuracy, a comparison study of these different models is conducted. The EfficientNetB0 model has achieved the highest test accuracy around 96.50% for Pneumonia X-ray images. The same work has been applied to the COVID-19 CT scans dataset, and the highest accuracy is achieved with the ResNet-50 network (99.5%). Therefore, these two models will be used for rapid diagnosis and assist radiologists in the detection process precisely.