Abstract
Chest X-rays are valuable for diagnosing and monitoring various conditions such as pneumonia, COVID-19, and heart disease, to cite but a few. This study specifically focuses on X-ray image classification between non-pathological lung images (i.e., normal/healthy) and lung images affected by pneumonia. The latter category is further subdivided into common pneumonia (bacterial and viral) and COVID-19 pneumonia. In this paper, we experimentally compare state-of-the-art machine learning methods tackling this multi-class classification problem. Specifically, we design a comparative methodology to put to test six prominent Convolutional Neural Network architectures, including ResNet, AlexNet, VGGNet, SqueezeNet, DenseNet, and InceptionV3. Our work overcomes the issue that the existing proposals for tackling pneumonia classification based on Deep Learning are not directly comparable, due to considerable differences in their respective modeling pipeline. In this paper, we conduct a comparative evaluation based on a uniform pipeline, from data collection, curation, and all the way to model validation. We introduce a balanced, labeled dataset, including 6,939 X-ray images. In pursuit of high accuracy, we also integrate (in all six cases) the same data curation, transformation, and augmentation techniques. Key findings demonstrate VGGNet's superior performance, with an overall 97% accuracy, making it highly applicable to real-world clinical scenarios.