Benchmarking Transformer Architectures for Chest X-ray Classification

Lung diseases remain a major global health concern, necessitating accurate and timely diagnosis. Chest X-ray (CXR) imaging is widely used but challenging to interpret due to overlapping radiographic features and subjective variability among radiologists. Deep learning approaches, particularly Convolutional Neural Networks (CNNs), have shown promise but are limited in capturing global spatial dependencies. Vision Transformers (ViTs) overcome this limitation through self-attention, making them increasingly attractive for medical image analysis. This study systematically evaluates 13 Transformer-based architectures across three CXR datasets with distinct tasks: Pneumonia (3-class: Normal, Bacterial, Viral), COVID-QU-Ex (3-class: Normal, Non-COVID Pneumonia, COVID-19), and Tuberculosis (2-class: Normal, Tuberculosis). All models were trained under a unified setup with consistent preprocessing, augmentation, and evaluation protocols. To improve robustness, a soft voting ensemble of the top five models was also implemented. Results demonstrate that Transformer-based models provide highly competitive performance. On the Pneumonia dataset, the ensemble achieved an accuracy of 0.8743 and F1-score of 0.8615, surpassing several single models such as DeiT-Base (F1 = 0.8725). On COVID-QU-Ex, the ensemble soft voting obtained 0.9593 accuracy and 0.9582 F1-score, effectively balancing precision and recall. On Tuberculosis, ViT-B/16 and MobileViT-S achieved perfect performance (F1 = 1.0), likely influenced by dataset imbalance. These findings highlight the clinical potential of Transformer-based models, particularly when combined through ensembles, for robust and accurate CXR classification.