A Deep Learning Based Approach on CXR Images for Tuberculosis Detection Using Vision Transformer

Tuberculosis (TB) exists as a global healthcare issue especially in areas with limited medical facility and its early intervention is essential to prevent its adverse effects. While chest X-ray (CXR) imaging is widely used for TB screening, manual interpretation is often inconsistent and can lead to diagnostic delays. This paper proposes an automated TB detection framework using Vision Transformers (ViTs) to classify CXR images. Contrary to conventional convolution neural networks (CNNs), ViTs capitalizes on self-attention mechanisms to model global dependencies across the image, enhancing diagnostic performance. To further enhance model generalization, have utilized state of the are data augmentation a technique, like Mixup and CutMix, addressing common challenges like data imbalance. Contrast Limited Adaptive Histogram Equalization (CLAHE) is employed to enhance image clarity. The suggested model was evaluated with an accuracy, precision, recall & F1 score of 99.52%, 100%, 96.97%, and 98.46% respectively. The model's ROC score was 98.48%, indicating its strong ability to distinguish between TB and non-TB cases. These results demonstrate the potential of the ViT-based approach to improve TB diagnosis, offering a reliable and efficient tool for healthcare systems, particularly in regions with limited access to trained radiologists