A Multi-Stage Deep Learning Approach to Tuberculosis Detection with Explainable Insights

Accurate and timely diagnosis of Tuberculosis (TB) is crucial, given its continued impact on global health. While chest X-rays are a common diagnostic tool, they are susceptible to subjective interpretation by radiologists. To enhance the reliability of TB detection, we present an AI-powered solution that integrates segmentation, classification, and explainability. Our approach utilizes UNet++ for precise lung region segmentation, which improves the accuracy of subsequent classification. By leveraging EfficientNetB0, our model delivers outstanding accuracy rates: 99.76% on the Tuberculosis Chest X-ray Dataset, 93.4% on the Montgomery and Shenzhen datasets, and 97.4% when all datasets are combined. To further enhance transparency in decision-making, we incorporate Explainable AI (XAI) techniques, including LIME, SHAP, and Grad-CAM. Our approach is more accurate than current works and can be used on a large scale. Due to its precision and interpretability, we believe it can significantly improve TB detection worldwide and ultimately help more patients get the necessary care.