Comparative Analysis of CNN Models with Vision Transformer on Lung Infection Classification

Lung infections are the most frequently observed medical conditions that affect the respiratory system. These infections can have meager effects like cough and if ignored can lead to the cause of death. Hence, a classification model that helps in the early detection of lung infections can help in avoiding further complications. This paper focuses on techniques to classify lung infections based on different convolution neural network model architectures in comparison with advanced deep learning techniques like Vision Transformer trained on chest X-ray dataset. One of the most powerful imaging techniques for diagnosing lung infections is chest X-ray. So, our model is built on chest X-rays collected from 30,805 individuals that were classified into 15 labels (including no finding). The dataset consists of 112,120 samples of images that are utilized for developing the model for predicting and analyzing lung infections by using InceptionV3, ResNet50, VGG16, InceptionResNet, and Vision Transformer. Each model is evaluated on the basis of mean accuracy error (MAE), binary accuracy, and loss. Among all the models, InceptionResNet has obtained a best accuracy of 93.33%. This study signifies that Vision Transformers are yet to be developed in order to catch up with CNN models.