A Hybrid CNN-Transformer Framework for Automated Tuberculosis Detection from Chest X-rays

Abstract

Tuberculosis (TB) remains a major global health challenge, demanding rapid and accurate diagnosis to enable timely treatment and reduce disease transmission. Although chest radiography is widely used for TB screening, manual interpretation is often time-consuming, subjective, and constrained by the shortage of experienced radiologists, especially in high-burden, resource-constrained settings. To address these limitations, this study proposes HybridTBNet, a hybrid deep learning framework that integrates Convolutional Neural Networks (CNNs) and Transformers for automated TB detection from chest X-ray images. The model is designed to capture both local pathological patterns and global contextual dependencies through hierarchical feature extraction and attention-based representation learning. A curated multi-source dataset of 6,708 high-quality chest X-ray images, comprising 3,194 TB and 3,514 Normal cases after quality filtering, was used for development and evaluation. A strict patient-wise split was applied to prevent data leakage, resulting in separate training, validation, and held-out test sets; however, minor class imbalance persisted in the final test subset. On the unseen test set, HybridTBNet achieved 98.2% accuracy, 97.1% sensitivity, 98.5% specificity, 94.9% F1-score, 0.991 ROC-AUC, and 0.978 PR-AUC, demonstrating strong potential as an effective automated TB screening support tool.