Identification of Pneumonia with X-ray Images Using Deep Transfer Learning

Abstract

Radiology plays an essential role in the identification of pathologies; however, image interpretation and the guarantee of accurate diagnoses continue to represent a challenge that involves expert radiologists. This study proposes a model to identify pneumonia in chest X-rays using the deep transfer learning technique, where five pre-trained network architectures and a classifier are tested. The images used in this work are categorized into bacterial pneumonia, viral pneumonia, and normal cases. The pre-trained models used include DenseNet201, MobileNet, VGG16, VGG19 and ResNet50. A support vector machine is used as a classifier. Results show that the ResNet50 model performs best in binary classification (pneumonia vs. non-pneumonia) with 98.1% accuracy and 98.7 F1-score. For multiclassification, VGG19 performs best with an accuracy of 84.7% and an average F1-score of 81.1%. The methodology employed proved to be competent and outstanding when compared to other studies in the state of the art.