Abstract

Introduction: Tuberculosis (TB) remains a significant global health challenge, with high morbidity and mortality, particularly in low- and middle-income countries (LMICs). Early detection of TB is crucial for effective disease control and management. However, manual interpretation of X-rays can be subjective and resource-intensive. Machine learning (ML) and deep learning (DL) models have demonstrated potential in automating medical image analysis. This study compares the diagnostic performance of traditional ML models (Random Forest, Support Vector Machine, and k-Nearest Neighbors) and a deep learning model (DenseNet-169) for TB detection in chest X-ray images.

Methods: A comparative design was used to train and evaluate Random Forest, Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and DenseNet-169 models using chest X-ray datasets from Montgomery County and Shenzhen. The models were assessed using key performance metrics, including accuracy, precision, recall, F1-score, and ROC-AUC. Standard preprocessing techniques, such as normalization and data augmentation, were applied to ensure consistency across models.

Results: Random Forest outperformed the other models with an accuracy of 83.0%, precision of 86.7%, and recall of 76.5%, demonstrating a lower false negative rate compared to SVM and k-NN, both of which exhibited recall rates of 69.1%. The DenseNet-169 model achieved an accuracy of 74.5%, precision of 80.0%, and recall of 63.8%. Despite its strong precision, DenseNet-169’s lower recall rate suggests a higher risk of missed TB-positive cases, which is critical in public health contexts.

Conclusion: The study highlights that Random Forest is a reliable ML model for TB detection in chest X-rays, particularly in LMICs with limited specialized medical personnel. While DenseNet-169 showed potential, its lower recall rate raises concerns about its standalone use for TB diagnostics. The findings emphasize the need for balancing sensitivity and specificity in diagnostic models and suggest further research into hybrid approaches combining traditional ML and deep learning techniques. Validating models on diverse datasets and improving recall through techniques like transfer learning could further enhance TB diagnostic performance.