Vision Transformers Excel in Chest X-Ray Analysis

In this study, we evaluate multiple benchmark model architectures, including Vision Transformers (ViT), for multi-label classification of chest X-ray (CXR) images using the NIH CXR 14 dataset ─a challenging benchmark for medical image analysis with annotated radiographs covering a wide range of diseases. Our results show that ViT outperforms DenseNet in terms of accuracy and Area Under the Curve (AUC) for most pathologies. For example, the AUC for Cardiomegaly is 0.9150 with ViT compared to 0.9126 with DenseNet, and for Emphysema, the AUC is 0.9380 with ViT versus 0.9360 with DenseNet. This work highlights the tradeoffs between the localized feature extraction of Convolutional Neural Networks (CNNs) and the global contextual modeling of transformers, demonstrating the potential of both architectures in medical imaging tasks.