CXR-ViT: A Transformer-Based Framework for Enhanced Chest X-Ray Classification

Increasing need for accurate medical image analysis, chest X-ray classification has become a powerful tool for diagnosing pulmonary diseases. However, the inherent complexity of thoracic abnormalities and the limited annotated medical data pose significant challenges to accurate classification. In this study, we utilized fine-tuned deep learning-based CNN and large vision transformer frameworks to classify lung diseases from CXR images. Moreover, to address the issues of model interpretability and decision-making, we use Gradient-weighted Class Activation Map (Grad-CAM) to visualize the results and improve the model predictions’ interpretability and accuracy by identifying relevant regions in the lung areas. The results of our experiments on the publicly available COVID-QU-Ex dataset show that the proposed CXR-ViT outperforms the conventional CNN approaches with an accuracy of 99.01 %. This work highlights the use of transformers in medical image analysis, emphasizing better lung disease classification and interpretability.