A Hybrid Deep Learning Framework with Attention Mechanism for Tuberculosis Detection in Chest Radiographs

Abstract

Tuberculosis (TB) continues to be an important global health problem, despite the availability of effective treatment. TB is a particular challenge in resource-poor settings where access to expert radiological interpretation is limited. Here, we propose an attention-enhanced multi-backbone deep learning model for automatic TB classification using chest X-ray (CXR) images. The proposed architecture is a combination of the three pretrained convolutional neural networks (CNNs), EfficientNet-B3, ResNet152, and DenseNet169, to extract diverse and complementary features. These features are fused and refined by a lightweight attention module, which adaptively highlights the most discriminative representations. Finally, a multi-layer fully connected network does the classification. The model was trained and tested on a curated TB dataset and obtained a test accuracy of 0.991, an F1 score of 0.985, and an AUC of 0.988. These results show the robustness and efficiency of the model for detecting TB-positive cases from normal CXR images. We incorporate the attention mechanism and multiple backbones to make the features more expressive and improve generalization ability to different qualities of images. This approach offers a scalable and reliable solution for TB screening with the potential to support clinical decision-making and reduce diagnostic burden in high-prevalence regions.