Interpretable Deep Learning Model for Tuberculosis Detection Using X-Ray Images

Tuberculosis (TB) is a worldwide severe health concern that causes numerous deaths yearly. Detecting TB promptly and precisely is crucial for limiting its effects and preventing fatalities. Chest X-ray (CXR) images play a critical role in TB diagnosis, but identifying the disease in its early stages is complex and can result in lengthy and expensive treatments. This study introduces a novel approach for TB detection by implementing a lightweight parallel CNN model (LP-CNN) with fewer parameters and computational requirements. This model boasts exceptional efficiency in distinguishing between normal, non-TB (bacterial and viral pneumonia), and TB cases. To discover the model’s comprehension, we employ the Shapley Additive Explanations (SHAP) direction to clarify the prediction process. The transfer learning (TL) models evaluated include DenseNet169, DenseNet201, InceptionResNetV2, VGG19, and Xception, and their performances are evaluated based on accuracy, precision, recall, F1-score, and AUC metrics. The LP-CNN model shows outstanding performance, achieving high scores in precision (99 ± 1.36E-16%), recall (98.33 ± 0.0115%), F1-score (99.33 ± 0.0057%), accuracy (99.46%), and AUC (99.93%). It is worth noting that LP-CNN has a compact design, consisting of only 1.658 million parameters, eight convolutional layers (CL), and a size of 19.1 MB. The results of this research provide an effective way to analyze chest X-rays to detect tuberculosis quickly and accurately. With the higher accuracy of the LP-CNN model and the interpretability of SHAP, there is significant potential to improve TB diagnosis and thereby reduce consequences and disease spread. This outcome highlights the revolutionary effect of AI-driven diagnostics in practical tuberculosis analysis, providing a potential step toward efficient disease treatment without radiologists.