A comparison of deep learning models for pneumonia detection from chest x-ray images

Abstract

Purpose: The aim of this study is to develop an automatic pneumonia detection system for disease detection by effectively extracting disease-related features from chest X-ray images.Theory and Methods: Three different deep learning approaches are proposed for automatic detection of pneumonia. These approaches are deep feature extraction, transfer learning, and end-to-end learning. In experimental studies, 10 different pre-trained convolutional neural network models (AlexNet, VGG16, VGG19, ResNet50, DenseNet201, DarkNet53, ShuffleNet, SqueezeNet, MobileNetV2 and NasNetMobile) were used and a new network was trained from scratch. The extracted features are classified with the support vector machine, k nearest neighbor and random forest classifiers.Results: The success of the fine-tuned AlexNet model produced an accuracy score of a 98.50%, which was the highest of all results achieved. In the deep feature extraction method, the ShuffleNet model showed the highest success rate of 98.00% among all models. The end-to-end training of the developed CNN model yielded 96.75% results.Conclusions: As a result, in this paper, a new chest X-ray pneumonia dataset is introduced. Various deep learning approaches are employed for pneumonia detection on this new dataset.