Deden Witarsyah, Hadi Almohab, Riswan Septriayadi Sianturi, Dedy Syamsuar
Tuberculosis (TB) remains a critical global health challenge, with early detection often hindered by limited diagnostic resources and variability in manual Chest X-Ray (CXR) interpretation. This study evaluates the diagnostic performance and computational efficiency of four deep learning architectures—VGG16, DenseNet121, MobileNetV2, and a custom lightweight CNN—trained for 10 to 100 epochs on a multisource dataset comprising 4,200 CXR images. Stratified data splitting and data augmentation were employed to address class imbalance and improve generalization. Among the evaluated models, VGG16 achieved the highest classification performance, with 99.21% accuracy, an F1-score of 0.98, and a ROC-AUC of 0.99. In contrast, MobileNetV2 provided a favorable trade-off between accuracy (98.73%) and inference efficiency, achieving the fastest prediction time of 12 ms per image, supporting its potential use in resource-constrained environments. Although the results indicate that transfer learning with optimized training protocols can support accurate and efficient AI-assisted TB screening, this study is limited by the lack of patient-level data separation, statistical cross-validation, and model explainability. Future work will address these limitations by incorporating explainable AI methods, statistically robust validation strategies, and evaluation on independent and diverse clinical datasets. Copyright © 2026 by the authors
Faculty of Computer Science, Universiti Tun Hussein Onn Malaysia, Malaysia; Faculty of Computer Science, Universitas Brawijaya, Indonesia; Faculty of Engineering, Computer and Design, Nusa Putra University, Indonesia; School of Information Systems, Bina Nusantara University, Indonesia
