Refining the role of the chest radiograph in paediatric tuberculosis clinical care and research

The proportion of individuals that die of tuberculosis (TB) each year that are children is disproportionately higher than the proportion of incident TB that is seen in children. This is largely due to limitations in the diagnosis of disease, with an estimated 95% of children dying from TB who went undiagnosed. To close the case detection gap for paediatric TB and reduce unnecessary childhood deaths, especially in young children, better strategies are needed to diagnose TB and facilitate access to appropriate treatment. In clinical research, more robust strategies are also required to establish TB disease with certainty in children, to ensure that research methods are sound, and results are understandable and generalisable. Paediatric TB is usually clinically diagnosed in the absence of microbiological confirmation, and chest X-ray (CXR) is a critically important tool in diagnostic decision-making algorithms. In this PhD dissertation I evaluate how CXR, as a tool for both diagnosis and disease severity stratification, could be better used and its role refined, within paediatric TB clinical care and research. I also investigate the role of novel computer-aided detection (CAD) approaches to CXR in the landscape of paediatric TB diagnosis and management. In symptomatic children presenting to care with presumptive TB, and enrolled on a prospective TB diagnostic study, I show that certain individual CXR features have high specificity for microbiologically confirmed TB and are strongly associated with TB in a regression model. I also show that some CXR features, while commonly seen in paediatric TB, have less value in distinguishing TB from non-TB lung disease. Using the same dataset, I present evidence that CAD (Delft’s CAD4TB v7.0) performance to distinguish TB disease from non-TB disease is suboptimal in young children, but that its performance can be significantly improved after fine-tuning with well-characterized paediatric CXRs. I describe the current paediatric CAD landscape and define the role of the CAPTURE consortium project in closing some of the evidence and access gaps around TB-CAD for children. This includes the development of a large, well-curated repository of CXRs from children investigated for TB, independent evaluation of current CAD algorithms for performance in children (where I show that performance is poor) and CXR image sharing to promote the optimisation of these algorithms for use in children. For CXR used as a tool to stratify the severity of TB disease, I describe how the theoretical concept of paediatric TB disease severity was adapted into a practical, implementable approach to CXR classification in a large phase 3 TB treatment-shortening trial, “SHINE”. I show that this approach was feasible to implement in the trial, but that it was resource intense. I propose that CAD for disease severity stratification has the advantage of being standardised and not reliant on access to specialist clinicians and I explore CAD for this use-case. I show that, despite not being specifically trained for disease severity stratification or with paediatric images, CAD scores generated from two different CAD algorithms were significantly different between CXRs classified as severe versus nonsevere by human expert readers. Finally, I present several resources that I have developed, together with the International Union Against Tuberculosis and Lung Disease, to support training and capacity building for CXR interpretation by non-specialist healthcare workers in the context of paediatric TB seen in low-resource settings. Throughout the dissertation I highlight how each analysis has combined with other synergistic work to refine how CXRs are used to diagnose and manage paediatric TB better, in order to inform World Health Organization (WHO) guidelines and change clinical practice. The latest WHO guidelines for paediatric TB (2022) include evidence-based Treatment-Decision Algorithms for paediatric TB which include CXR scores that were developed from our data-driven prediction model. These guidelines also include treatment shortening guidance based on the disease severity stratification approach developed for the SHINE trial. The CAPTURE consortium project is ongoing, and the evidence generated will inform the WHO’s paediatric CAD evidence review planned for 2026. ‘Alone you will go fast, together you will go far.’ This body of work represents only a small component of the work that has been undertaken over the last decade by a dedicated group of paediatric TB clinicians and researchers who are trying to close the case detection gap in childhood TB. My contribution shows that we can use an existing and established tool, CXR, to develop data-driven approaches that improve the diagnosis and clinical management of TB in children. It also shows that old tools can be used in novel ways: CAD, which relies on CXR acquisition to be accessible, offers the potential to automate several decision-making steps in the paediatric TB management pathway.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Palmer, Megan

Advisors dc:contributor.advisor

• Seddon, James Alexander
• Hesseling, Anneke C.
• Goussard, Pierre

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/135962

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/135962