Hierarchical CNN Framework for Tuberculosis Detection UsingChest Radiographs

Abstract

Tuberculosis (TB) remains a leading cause of death worldwide, with over 10.6 million people falling ill and 1.3 million dying from the disease in 2022, according to the World Health Organization. Chest X-rays (CXR) are an essential diagnostic tool in early TB detection, but manual interpretation suffers from inconsistencies and resource limitations. Traditional diagnostic methods are prone to subjectivity and require skilled radiologists, leading to delayed or inaccurate TB identification, particularly in under-resourced regions. To address these challenges, this study proposes a robust and automated approach for tuberculosis classification using chest X-ray images, focusing on two diagnostic categories: Normal and Tuberculosis. The novelty of this work lies in its three-tiered contribution. First, a comprehensive image preprocessing pipeline is applied. Second, two conventional classifiers Decision Tree Classifier (DTC) and Random Forest Classifier (RFC) are implemented and evaluated as baseline models. Lastly, the core contribution is the development of a Hierarchical Feature-driven Convolutional Neural Network (HF-CNN), which leverages both low-level and high-level spatial features in a multi-layered learning framework. Unlike conventional CNNs, the proposed HF-CNN architecture integrates hierarchical feature aggregation mechanisms to capture subtle pathological patterns specific to TB, enhancing both sensitivity and specificity. This end-to-end framework not only automates the detection process but also significantly improves classification accuracy, demonstrating strong potential for deployment in clinical decision support systems.