Class-attention pooling and token sparsity based vision transformers for chest X-ray interpretation.

Chest radiography is the most widely used imaging technique for the diagnosing of lung diseases, but interpreting X-rays can be very difficult due to the subtle abnormalities and variations in image quality taken during X-rays. While deep learning models, especially the Convolution Neural Networks (CNNs), have shown strong performance when it comes to detecting pneumonia and other conditions, Vision Transformers (ViTs) have recently surpassed CNNs on several chest X-ray benchmarks by dividing images into small patches and learning global relationships. However, standard ViTs can sometimes focus on irrelevant regions, making their decisions less interpretable. To address this, we propose an enhanced ViT model tailored for chest X-ray analysis that prioritizes both accuracy and explainability. Our model introduces class-attention pooling technique, where each disease-specific class token learns to highlight relevant regions of the image, improving disease-wise focus. Token sparsity and random token dropping further help the model attend to only the most informative patches, enhancing robustness against noise. A convolutional stem is added before patch creation to extract fine local features like edges and textures, ensuring early capture of lung-specific patterns. Additionally, each X-ray undergoes preprocessing using Contrast Limited Adaptive Histogram Equalization (CLAHE), which enhances local contrast and makes subtle lesions more visible. The model is trained with mixed-precision computation, a warm-up cosine learning rate schedule, and the AdamW optimizer, allowing stable and efficient training on large datasets. It is then evaluated on Tuberculosis Chest X-Rays and Pulmonary Chest X-Rays datasets which are publicly available, the proposed framework achieved 99.19% Training accuracy, a Validation Accuracy of 97.78%, an F1-score of 0.94, and an AUC of 0.99, outperforming the baseline ViT. It is pointed out that the above scores are obtained based on the image-level split, owing to the limitations of the dataset, and the performance may be over-estimated compared to the validation on the patient level. The Grad-CAM heatmaps further confirm the fact that the model focuses on clinically relevant areas such as opacities or nodules, reinforcing interpretability and trust. Overall, this improved ViT framework offers both high diagnostic accuracy and also clear visual explanations, implying its possible usage in acting as an AI assistant for radiologists in efficiently detecting lung diseases.