Advancements in Deep Learning for Tuberculosis Detection: A Review of Modified CNN Architectures and Transfer Learning Approaches for Chest X-ray Analysis

Tuberculosis remains a significant global health challenge, claiming over 10 million lives annually. Chest radiography continues to be a pivotal tool in the detection and diagnosis of tuberculosis. However, the reliability of human interpretation is hindered by inter-observer variability. To address this, researchers have actively explored deep learning-powered algorithms to enhance the accuracy and efficiency of tuberculosis diagnosis from chest X-rays. This review examines recent research studies that employ deep learning models primarily consisting of modified convolution neural networks with advanced optimization techniques and transfer learning approaches to leverage pretrained networks effectively. Based on these studies, the deep learning models in tuberculosis detection from chest x-rays show promising ranges. Accuracy ranges from 77.14% to 99.6%, sensitivity from 67% to 100%, and specificity from 61% to 100%. F1-scores are between 50% and 99.29%; the area under the curve varies from 0.83 to 0.999, indicating strong model performance across studies. These metrics underscore the high diagnostic potential of deep learning models with modified architectures and transfer learning contributing to improved reliability and generalization in TB classification. This increased reliability will further assist the diagnosis of TB and mitigate the hurdles in treatment caused by the social and economic conditions of the patients.