Systematic evidence based review of deep learning methods for tuberculosis detection and classification from chest X ray images

Background: Despite developments in therapy and diagnostics, TB remains an enormous global health challenge, especially in resource-poor environments with restricted diagnostic ability. Chest X-ray (CXR) is a cost-effective and commonly used screening modality, but has limitations with regards to inter-observer variation and shortage of experts in the field of radiology for interpretation. Recent progress in deep learning, notably convolutional neural networks (CNNs), demonstrates good promise in automated TB detection. Objective: This review provides a comprehensive overview of the current published works of CNN-based methods for TB detection using CXR, and maps evidence on dataset, models, performance, and gaps, which could guide future work. Methods: In accordance with PRISMA standards, a thorough search was performed on ScienceDirect, PubMed, IEEE Xplore, MDPI, and ACM Digital Library. Only peer-reviewed journal papers using CNN-based deep learning techniques were considered. We further used various visualization techniques to reinforce our analysis: (i) a network visualization graph, depicting connections between highly-cited papers; (ii) data visualizations, presenting the datasets and data types (public vs. private, size, and imaging modality); (iii) a pie chart to visualize the distribution of the used CNN architectures; and (iv) summary tables, summarizing the main studies and including details on the datasets used, the models employed, performance scores, and limitations mentioned. Results: A representation of the prevalence of CNN (Specifically VGGNet, ResNet50, DenseNet, EfficientNet and Unet) models, and many works have achieved above 90% accuracy on TB detection. However, limitations are dependent on small or imbalanced datasets, low generalizability between populations, and no interpretability in AI predictions. Visualization analysis also identified that research evidence seemed to be clustered around publicly available datasets and an emerging interest in transfer learning approaches. Conclusion and Future Trends: CNN-based deep learning methods show promising results in automated detection of TB from chest radiographs. However, large-scale, diverse, and clinically validated datasets that will be suitable for multi-center testing and explainable AI frameworks and cross-domain generalization are still key challenges. In the future, it will be of interest to explore lightweight architecture for optimal deployment in low-resource-level environments, multimodal fusion with clinical data, and advanced algorithms such as federated learning to improve the scalability and transferability in clinical scenarios.