AI brings TB screening four kilometres underground at South Africa’s Mponeng mine

Four kilometres underground at South Africa’s Mponeng Gold Mine, portable X-ray kits and offline AI are bringing tuberculosis screening to the workplace. Google Research says miners can receive an initial assessment within seconds and, if a scan suggests TB, be referred immediately for confirmatory molecular testing. The advance is a shorter route into care for people whose work puts a radiology department a long journey away.[1]

Google’s 23 October account reports more than 25,000 chest radiographs assessed by over 90 clinicians across approximately 40 facilities. The programme covers South Africa, Sierra Leone and Côte d’Ivoire, with expansion under way in the Democratic Republic of the Congo, Zambia and Cameroon. The company reports turnaround times below 45 seconds; the figures describe the wider programme, not the number of miners screened at Mponeng.[1]

What happens between the image and the treatment

A chest radiograph can suggest pulmonary tuberculosis, but a scan cannot by itself determine that someone carries the organism that causes the disease. Pneumonia, old scarring and other lung disorders can produce overlapping appearances. In the reported workflow, the AI looks for patterns suggestive of TB or other abnormalities, clinicians interpret the findings, and anyone requiring follow-up is referred for a confirmatory molecular test. It is the completion of that chain—from screening to reliable diagnosis to appropriate treatment—that matters for health.[1][2]

At Mponeng, chronic silica-dust exposure makes lung screening particularly relevant. Nexus AI worked with the Stop TB Partnership, Harmony Gold and South Africa’s Department of Health to take ultra-portable imaging equipment and offline inference underground. Google describes a pathway in which screening, triage and confirmatory testing can happen in one visit. That removes a hand-off at which a patient might otherwise need to travel, arrange another appointment or wait for a specialist report.[1]

Nexus AI has also provided more than 3,000 free screenings to under-resourced public clinics, according to Google, against a goal of 10,000 sponsored screenings annually. Screening volume measures access to the service. Following confirmed cases through treatment would show how much that improved access changes health outcomes.[1]

The work required to reach a clinic

Google traces the technology from earlier chest-imaging research through external evaluation, prospective validation and integration by Nexus AI. According to its account, the partner obtained CE Medical Device Regulation class IIb certification and ISO 13485 certification for quality management. It is also pursuing World Health Organization prequalification through an agreement with FIND. These steps matter because real clinics require traceability, equipment compatibility, maintenance and local regulatory oversight. A high classification score on a research dataset cannot substitute for those operational duties.[1]

Alongside the Nexus deployment, Google is supplying adaptable medical models for other health systems. MedSigLIP links medical images with clinical language, while MedGemma combines imaging and clinical information. Indonesia’s Ministry of Health is fine-tuning them on domestic radiographs for its screening initiative. The common design choice is local adaptation: a useful model must fit the images, equipment and clinical pathway where it will run.[1]

The full economics remain difficult. A portable X-ray machine costs money to acquire and maintain; a trained operator must be available; people flagged by software need onward care; clinics need power, infection control and reliable medicine supplies. Moving the first assessment from weeks to seconds does not remove those constraints. It can, however, stop a scarce radiologist's availability from becoming the bottleneck before anyone knows whether follow-up is warranted.[1]

A shorter route into care

The underground deployment makes the value of local inference tangible. A network connection and a distant specialist no longer have to stand between taking an image and deciding who needs closer attention. The clinical value still depends on completing the next steps, but the first step has moved much closer to the people who need it.[1]

Report an error or suggest a correction