Comparative study of CNN techniques for tuberculosis detection using chest X-ray images from Indonesia

Abstrak

Convolutional neural networks (CNNs) represent a popular deep-learning approach for image classification tasks. They have been extensively employed in studies aimed at classifying tuberculosis (TB), coronavirus disease 2019 (COVID-19), and normal conditions on chest X-ray images. However, there is limited research utilizing Indonesian data, and the integration of CNN models into user-friendly interfaces accessible to healthcare professionals remains uncommon. This study addresses these gaps by employing three CNN architectures—AlexNet, LeNet, and a modified model—to classify TB, COVID-19, and normal condition images. Training data were sourced from both a local hospital in Indonesia (RSUP dr. Rivai Abdullah) and an additional online dataset. Results indicate that AlexNet achieved the highest accuracy, with rates of 97.52%, 64.45%, and 92.43% on the Kaggle dataset, the RSUP Dr. Rivai Abdullah dataset, and the combined dataset, respectively. Subsequently, this model was integrated into a user interface and deployed for testing using new data from the RSUP Dr. Rivai Abdullah dataset. The web-based interface, powered by the Gradio library, successfully detected 7 out of 10 new cases with 70% accuracy. This implementation may enable medical professionals to make preliminary diagnoses.