CNN-Based Approaches for Tuberculosis Detection in Medical Imaging: A Comparative Analysis

Tuberculosis (TB) has a high mortality rate in many regions and, hence, is a critical infectious disease. The World Health Organisation’s (WHO) 2018 Global TB study reports that approximately 1.5 million people die from TB annually, and around 10 million people are infected with TB each year. Studies indicate that TB kills more than 4,000 individuals daily; early diagnosis of the illness might have helped to avoid this death toll. Using advanced technology, automatically analyzing and categorizing chest X-rays into TB and non-TB groups provides a dependable alternative to the personal judgments usually made by healthcare professionals. We present the performance stats of four Convolutional Neural Network (CNN) models—VGG16, UNet, ResNet50, and DenseNet169. We applied binary classification to the models and, after analyzing the result, noticed that ResNet50 performed exceptionally well in classification tasks. Whereas U-Net outperformed it in segmentation tasks. Although VGG16 delivered decent results with its more straightforward design, DenseNet169 demonstrated good recall but had lower precision. This approach improves diagnostic precision, lowers the possibility of human error, and provides a scalable option for tuberculosis detection—especially in regions where access to qualified radiologists is scarce. With each model’s mentioned strengths and weaknesses, we can easily understand their suitability for real-world medical imaging applications. This knowledge can guide future research and practical use in diagnosing TB and other lung diseases through medical image analysis.