In the radiology reading room of a community health centre in Khayelitsha, the screen never goes dark. By nine on a typical Tuesday morning, Nomvula Dlamini, the facility's sole radiographer, will have reviewed sixty-three chest X-rays. By afternoon, the queue will have grown. The Western Cape's public health system was designed, at its most generous resourcing, for a patient load approximately forty percent lower than what it now handles daily. South Africa diagnoses roughly 300,000 new tuberculosis cases each year — among the highest rates per capita on the planet — and the bottleneck is almost always the same: a shortage of radiologists to read the images that determine a patient's diagnosis and treatment pathway.
Into this constraint, a Cape Town-based startup called Lumis Health has inserted a machine-learning platform that reads chest X-rays for TB indicators with a sensitivity and specificity that, in independent clinical trials conducted at Tygerberg Hospital in 2024, matched or exceeded the performance of experienced radiologists on a dataset of 14,000 images. The system — built on a convolutional neural network trained on over 900,000 annotated images from across sub-Saharan Africa — can return a reading in under ninety seconds, flags cases by severity, and integrates directly with the National Health Laboratory Service's existing digital infrastructure. Dr Amahle Khumalo, Lumis's clinical director, spent four years in rural Limpopo before joining the company. She understands, viscerally, what delay costs.
Khumalo navigates a live-demo dashboard from the company's offices in Salt River, showing the previous day's queue from one of their deployment sites in the Eastern Cape. A high-probability TB flag appears on screen: a faint opacity in the right upper lobe, easily missed on a pressured morning. "A radiographer alone might have rated that as low-priority and it goes into the pile," she says. "The model flags it as urgent. That patient gets onto treatment within forty-eight hours instead of potentially three weeks." At scale, the difference is survival versus mortality. The Eastern Cape site processed 4,200 X-rays in March using the platform. Three radiologists doing that work manually would take eleven weeks.
The model flags it as urgent. That patient gets onto treatment within forty-eight hours instead of potentially three weeks.
The platform is not without its complications. A 2025 peer review published in The Lancet Digital Health noted that the model's performance on paediatric X-rays — a smaller but clinically critical subset — was measurably lower than on adult images, a finding Khumalo describes as the company's primary current research priority. There is also the question of clinical governance: who carries liability when an AI system flags an image incorrectly, and how radiographers are trained to treat the model's output as a tool rather than a verdict. "We are very clear that this is decision support, not decision replacement," Khumalo says. The Western Cape Department of Health has appointed a dedicated clinical oversight committee specifically for this purpose.
The global health technology sector has taken notice. The World Health Organization's Global TB Programme included Lumis's deployment model in a 2025 working paper on AI diagnostics in high-burden countries, the first time a South African company has featured in a WHO TB guidance document. Separately, the Gates Foundation has committed preliminary funding to a feasibility study extending the platform to three countries in East Africa, contingent on a Phase 3 efficacy trial currently recruiting participants across South Africa's public health system. For a startup that had seven employees and no revenue twenty-six months ago, the trajectory is steep.
Late afternoon in the Khayelitsha reading room, the queue has cleared. Nomvula Dlamini reviews the system's flagged cases for the day — seventeen high-priority, four referred immediately to the TB clinical team — and signs off. The platform handled 200 images. She reviewed 63 herself. Together, they processed what a single radiographer working an unassisted twelve-hour shift could not have touched. It is an unglamorous collaboration, transactional and quiet, running in a health centre built for half the load it carries — but functional in a way that health technology in this context rarely manages to be: calibrated to the actual conditions of the system it was designed to serve.
