A Short Review on Technological Advances in Tuberculosis Diagnosis

Background and Aims: Tuberculosis remains a critical global health challenge, particularly with the increasing prevalence of drug-resistant strains that undermine conventional treatment protocols. Traditional diagnostic approaches suffer from limitations in sensitivity, speed, and accessibility, especially in resource-constrained regions. This review aims to evaluate how computational intelligence technologies, particularly machine learning and deep learning methodologies, can revolutionize tuberculosis detection and management protocols to improve patient outcomes globally. Methods: This review examines recent progress in computational intelligence for tuberculosis diagnosis, emphasizing machine learning and deep learning. Key techniques include convolutional neural networks (CNNs) like VGG۱۶, ResNet۵۰, EfficientNetB۳, and Vision Transformers for chest X-ray analysis, and U-Net for detecting pulmonary issues. It also discusses machine learning methods such as Support Vector Machines (SVMs), Naive Bayes, and Decision Trees for imaging and non-imaging data. Validation often uses dataset splitting and cross-dataset testing with public datasets like Shenzhen (۶۶۲ CXRs) and Montgomery County (۱۳۸ CXRs) for generalizability. Innovative methods like cough audio analysis with Capsule Networks and Fully Convolutional Neural Networks (FCNNs) are highlighted for resource-limited environments. Results: The reviewed studies show significant improvements in tuberculosis detection using deep learning, with accuracies from ۹۳.۵۹% to ۹۹.۱% and sensitivity up to ۹۸%. A CNN-based CAD system achieved ۹۸.۴۶% accuracy on the Shenzhen dataset, while EfficientNetB۳ with U-Net segmentation reached ۹۹.۱% accuracy and ۹۹.۹% AUC-ROC. Image segmentation improved diagnostic precision by isolating pulmonary abnormalities. Cough audio analysis, a non-invasive method, attained ۹۷% accuracy, ۹۸% sensitivity, and ۹۶% specificity. Challenges include data quality issues, high computational demands, and integration into healthcare systems. Conclusion: Computational intelligence approaches offer transformative potential for tuberculosis management through enhanced early detection capabilities, personalized treatment regimens, and improved resource allocation. However, successful implementation requires addressing significant barriers, including the development of representative datasets, ethical considerations regarding algorithmic bias, and accessibility concerns for underserved populations. Future research should focus on creating lightweight models suitable for deployment in resource-limited settings and establishing standardized evaluation frameworks to facilitate clinical adoption.

کلیدواژه ها:

Tuberculosis ، Tuberculosis Diagnosis ، Computer-Aided Diagnosis ، Medical Imaging ، Deep Learning ، Machine Learning