AI X-ray Test Distinguishes TB, Pneumonia, Normal and Non-X-ray Images Quickly

Pneumonia and tuberculosis remain major global killers, and both are difficult to diagnose accurately from chest images. Pneumonia, most often caused by Streptococcus pneumoniae, accounts for 14% of deaths among children under five, causing 740,180 fatalities each year. Tuberculosis, caused by Mycobacterium tuberculosis, led to 1.25 million deaths in 2022, including 161,000 among people with HIV. Misdiagnosis is a serious problem: about 22.3% of pneumonia cases are misidentified as tuberculosis. To address these intertwined challenges, Rifatul Islam Majumder and collaborators developed a new image classification framework aimed at reading chest X-rays (CXR). The system is built to sort images into four categories — normal, pneumonia, tuberculosis and non-X-ray — the last of which is intended to catch outliers and images that do not belong to standard X-ray data. By designing the model to recognize non-X-ray inputs explicitly, the team aimed to improve the reliability of automated reads and reduce harmful misclassification between pneumonia and TB, which can lead to inappropriate care decisions.

The researchers trained convolutional neural networks using established architectures: a pre-trained ResNet-18 and a fine-tuned DenseNet-121. Models were trained both with and without a weighted loss function to test how class imbalances might affect performance. Training used a curated dataset drawn from multiple valid sources, containing 5,489 normal images, 4,273 pneumonia images, 4,197 tuberculosis images and 1,357 non-X-ray images. The testing reported very high overall performance: the best model reached 98.76% accuracy, with 99.01% precision and 99.03% recall, and it maintained or exceeded class-wise performance across the four categories. Including the non-X-ray class helped the framework detect outliers and unseen anomalies, increasing robustness. The use of popular, well-understood architectures like ResNet-18 and DenseNet-121 suggests the approach can be reproduced and adapted in a variety of settings.

The immediate implication of this work is a practical tool that could reduce diagnostic errors between pneumonia and tuberculosis and flag images that do not belong to standard X-ray categories. Because the system uses pre-trained and fine-tuned models rather than bespoke, resource-intensive pipelines, it is presented as a solution that can operate with minimal resources while aiming for maximum accuracy. For clinics and health systems in low-resource environments, an accurate automated read could speed diagnosis, guide proper treatment decisions, and help prioritize patients for further testing. By explicitly addressing the problem of outliers through a non-X-ray class and demonstrating high precision and recall, the framework points toward safer deployment of AI-assisted radiology tools. If validated in broader clinical trials and integrated responsibly, this kind of model could be a practical step toward better detection and management of pneumonia and tuberculosis globally.