Enhanced tuberculosis detection using Vision Transformers and explainable AI with a Grad-CAM approach on chest X-rays

• Published in: BMC medical imaging Vol. 25; no. 1; pp. 96 - 16
• Main Authors: Vanitha, K., Mahesh, T. R., Kumar, V. Vinoth, Guluwadi, Suresh
• Format: Journal Article
• Language: English
• Published: London BioMed Central 24.03.2025; BioMed Central Ltd; Springer Nature B.V; BMC
• Subjects: Accuracy; Algorithms; Artificial neural networks; Batch processing; Chest; Chest X-rays; Computer-aided manufacturing; Datasets; Decision making; Deep learning; Design; Electric transformers; Explainable AI; Explainable artificial intelligence; Global health; Grad-CAM; Humans; Imaging; Machine learning; Medical imaging; Medical imaging equipment; Medicine; Medicine & Public Health; Neural networks; Neural Networks, Computer; Public health; Radiographic Image Interpretation, Computer-Assisted - methods; Radiography, Thoracic - methods; Radiology; Recall; Self-attention; Tuberculosis; Tuberculosis - diagnostic imaging; Tuberculosis detection; Tuberculosis, Pulmonary - diagnostic imaging; Vision; Vision Transformer; X-rays; Deep learning; Tuberculosis detection; Explainable AI; Medical imaging; Chest X-rays; Grad-CAM; Diagnostic accuracy; Self-attention; Vision Transformer; Convolutional neural networks
• ISSN: 1471-2342; 1471-2342
• DOI: 10.1186/s12880-025-01630-3

Tuberculosis (TB), caused by Mycobacterium tuberculosis, remains a leading global health challenge, especially in low-resource settings. Accurate diagnosis from chest X-rays is critical yet challenging due to subtle manifestations of TB, particularly in its early stages. Traditional computational methods, primarily using basic convolutional neural networks (CNNs), often require extensive pre-processing and struggle with generalizability across diverse clinical environments. This study introduces a novel Vision Transformer (ViT) model augmented with Gradient-weighted Class Activation Mapping (Grad-CAM) to enhance both diagnostic accuracy and interpretability. The ViT model utilizes self-attention mechanisms to extract long-range dependencies and complex patterns directly from the raw pixel information, whereas Grad-CAM offers visual explanations of model decisions about highlighting significant regions in the X-rays. The model contains a Conv2D stem for initial feature extraction, followed by many transformer encoder blocks, thereby significantly boosting its ability to learn discriminative features without any pre-processing. Performance testing on a validation set had an accuracy of 0.97, recall of 0.99, and F1-score of 0.98 for TB patients. On the test set, the model has accuracy of 0.98, recall of 0.97, and F1-score of 0.98, which is better than existing methods. The addition of Grad-CAM visuals not only improves the transparency of the model but also assists radiologists in assessing and verifying AI-driven diagnoses. These results demonstrate the model’s higher diagnostic precision and potential for clinical application in real-world settings, providing a massive improvement in the automated detection of TB.