Science Portal
v2.1.prod.produseast1
Copyright © Médecins Sans Frontières
v2.1.prod.produseast1
BACKGROUND
The diagnosis of tuberculosis (TB) is difficult in children. Chest X-ray (CXR) plays a critical role to support TB diagnosis, but countries with the highest TB prevalence
often have the lowest availability of expert CXR readers, such as radiologists. AI-based computer-aided detection (CAD) technology is endorsed by WHO for CXR interpretation for TB in adults, but is not currently recommended in children. Because the CXR appearance of TB is different in children compared to adults, current CAD models that have been trained on adult datasets may not perform as well in children.
METHODS
This diagnostic study was nested in the TB ALGO PED study. The study cohort included 665 children under 10 years with symptoms of TB from study sites in Guinea, Niger, Nigeria and Uganda. All children had a frontal CXR which was analysed by CAD software trained and optimized for paediatric TB (qXR version 4.2.1, Qure.ai). The CAD results were compared to a reference standard of a consensus of three expert radiologist interpretations. To evaluate CAD performance, receiver operating characteristic (ROC) curves were generated and the area under the curve (AUC) was calculated overall and for each site.
RESULTS
The AUC of CAD for the entire study cohort was 0.76 (95% CI: 0.72-0.81), with AUC above 0.80 generally considered clinically useful. The AUC for Guinea, Niger, Nigeria and Uganda were 0.68 (0.42-0.94), 0.74 (0.68-0.80), 0.80 (0.70-0.90) and 0.87 (0.81-0.94), respectively. CAD performed better in sites with better image quality (technical factors and format).
CONCLUSION
CAD shows fair performance for TB detection on CXR in children, compared to expert radiologist interpretation, and is not yet ready for clinical use. Results are promising and CAD models will continue to improve with increasing access to paediatric CXR datasets. With improvements, CAD could have a significant role to increase TB diagnosis in children.