IMPACT-TB: Integrated Medical Imaging and AI for Precise Tuberculosis Detection

Abstract

Tuberculosis (TB) remains one of the most significant global health challenges, particularly in low-resource settings where timely and accurate diagnosis is critical for effective treatment and disease control. Despite advancements in diagnostic technologies, existing models often face limitations, such as high computational demands, limited generalizability across diverse populations, and challenges in interpretability. These constraints can hinder the widespread adoption of automated TB diagnosis systems, particularly in areas where the disease burden is high. To address these challenges, we propose IMPACT-TB, an advanced deep learning-based model that integrates cutting-edge feature extraction using the CoAtNet architecture and a robust fully connected neural network (FCNN) for precise TB diagnosis. The key steps of IMPACT-TB include detailed feature extraction from chest X-rays and CT scans using CoAtNet, followed by accurate classification through the FCNN, ensuring effective handling of complex and nonlinear relationships. The model is rigorously tested across multiple datasets, including TB-DSI, TB-DSII, and TB-DSIII, demonstrating consistent and superior performance with high accuracy of 96.56% for TB-DSI, 97.45% for TB-DSII, and 96.12% for TB-DSIII, respectively. Compared to existing models, IMPACT-TB not only achieves better diagnostic accuracy but also offers enhanced interpretability and generalizability, making it a valuable tool for TB diagnosis in diverse clinical settings.