Deep Learning Applications for Tuberculosis Detection on Chest X-rays in High-Burden Resource-Limited Settings: A Systematic Review with Implications for Africa
1
Faculty of Clinical Sciences, University of NIgeria Teaching Hospital, Enugu, Nigeria
Popul. Med. 2026;8(Supplement Supplement 1):
ABSTRACT
INTRODUCTION:
tuberculosis (TB) remains a major problem globally especially in high-burden resource-limited regions such as Sub-Saharan Africa, where weak infrastructure and HIV co-infection worsen diagnostic delays. Deep learning (DL) models on chest x-rays (CXR) provide a scalable, cost-effective, rapid screening tool. This systematic review assesses the DL performance for TB detection in such settings with specific relevance to Africa.
METHODS:
adhering to PRISMA guidelines, we identified primary studies from 2020-2025 via PubMed and DOI search. Relevant studies applied DL to TB diagnosis and using CXR and employed retrospective or diagnostic accuracy designs relevant to resource-limited contexts such as Africa. Extracted data included study design, population including HIVsubgroup, sample size 138-165754 CXR, model architectures (CNNs like ResNet and VGG usually pre-trained on ImageNet), datasets (e.g. NIH Shenzhen, South African surveys), validation strategies (cross validation external testing), performance metrics (AUC, sensitivity, specificity) and PROBAST risk-of-bias assessments. Quality was assessed for participant selection, index test, reference standards (e.g., xpert mtb/rif, culture), and analysis.
RESULTS:
Ten studies were included, reporting data from South Africa, Zambia, China, the US and analogous high burden regions. DL models achieved AUC 0.81-0.99, sensitivity 67-98.89% and specificity 79-100%, often matching or exceeding radiologist performance. High-yield studies including Kazemzadeh et al. and Rajpurkar et al., demonstrated noninferiority in HIV cohorts and cost reductions (40-80%). Explainability via Grad-CAM and saliency maps aided localization. PROBAST assessments revealed a low to moderate risk of bias, predominantly due to retrospective nature and generalizability issues.
CONCLUSIONS:
DL can facilitate TB diagnostics in African settings, supporting triage and equity. However, prospective validation, integration with clinical workflows, and addressing data biases are essential for widespread adoption.
tuberculosis (TB) remains a major problem globally especially in high-burden resource-limited regions such as Sub-Saharan Africa, where weak infrastructure and HIV co-infection worsen diagnostic delays. Deep learning (DL) models on chest x-rays (CXR) provide a scalable, cost-effective, rapid screening tool. This systematic review assesses the DL performance for TB detection in such settings with specific relevance to Africa.
METHODS:
adhering to PRISMA guidelines, we identified primary studies from 2020-2025 via PubMed and DOI search. Relevant studies applied DL to TB diagnosis and using CXR and employed retrospective or diagnostic accuracy designs relevant to resource-limited contexts such as Africa. Extracted data included study design, population including HIVsubgroup, sample size 138-165754 CXR, model architectures (CNNs like ResNet and VGG usually pre-trained on ImageNet), datasets (e.g. NIH Shenzhen, South African surveys), validation strategies (cross validation external testing), performance metrics (AUC, sensitivity, specificity) and PROBAST risk-of-bias assessments. Quality was assessed for participant selection, index test, reference standards (e.g., xpert mtb/rif, culture), and analysis.
RESULTS:
Ten studies were included, reporting data from South Africa, Zambia, China, the US and analogous high burden regions. DL models achieved AUC 0.81-0.99, sensitivity 67-98.89% and specificity 79-100%, often matching or exceeding radiologist performance. High-yield studies including Kazemzadeh et al. and Rajpurkar et al., demonstrated noninferiority in HIV cohorts and cost reductions (40-80%). Explainability via Grad-CAM and saliency maps aided localization. PROBAST assessments revealed a low to moderate risk of bias, predominantly due to retrospective nature and generalizability issues.
CONCLUSIONS:
DL can facilitate TB diagnostics in African settings, supporting triage and equity. However, prospective validation, integration with clinical workflows, and addressing data biases are essential for widespread adoption.
| eISSN: | 2654-1459 |
We process personal data collected when visiting the website. The function of obtaining information about users and their behavior is carried out by voluntarily entered information in forms and saving cookies in end devices. Data, including cookies, are used to provide services, improve the user experience and to analyze the traffic in accordance with the Privacy policy. Data are also collected and processed by Google Analytics tool (more).
You can change cookies settings in your browser. Restricted use of cookies in the browser configuration may affect some functionalities of the website.
You can change cookies settings in your browser. Restricted use of cookies in the browser configuration may affect some functionalities of the website.
