Portable X-rays and offline AI bring TB screening to South Africa’s deepest mine

Portable X-ray equipment and offline AI are bringing tuberculosis screening to miners four kilometres underground at South Africa’s Mponeng gold mine. Google Research describes a deployment by Nexus AI in which a suspicious scan can trigger immediate referral for confirmatory molecular testing. Screening, triage and confirmation can be organised within one visit, shortening a pathway that often requires repeated journeys.[1]

Google’s 23 October update reports that Nexus’s wider rollout spans approximately 40 healthcare facilities, where more than 90 clinicians have assessed over 25,000 chest radiographs. The company says turnaround has fallen from weeks to under 45 seconds. Those totals cover the broader programme, not the mine alone; they describe screening activity and speed rather than treatment outcomes.[2][1]

Why the location changes the clinical problem

Tuberculosis is treatable, but screening programmes need more than effective drugs. They need a way to identify people who ought to have diagnostic testing, and they need to do it where exposure and access problems are greatest. Mining creates an especially demanding setting: silica dust exposure and crowded occupational environments are important concerns, while a specialist radiologist may not be nearby. Moving a patient through several disconnected appointments can turn a technically available test into a practically inaccessible one.[1]

Chest radiography is useful for finding changes suggestive of pulmonary TB, but a suspicious image is not definitive proof of infection. The AI examines the X-ray for patterns that might warrant further investigation. A qualified clinical workflow must still review the result, decide whether a molecular test or another examination is required, and manage people with symptoms despite an unremarkable image. In Google's account, miners who screen positive can be referred for confirmatory molecular testing on the same visit. That last hand-off matters as much as the speed of image classification.[1]

The technology has a history longer than a single October announcement. Google researchers had previously evaluated deep-learning approaches for TB abnormalities on chest X-rays; partners subsequently worked through medical-device requirements and deployment questions. The recent update says Nexus achieved European CE MDR Class IIb certification for its product, together with ISO 13485 quality-management certification, and is pursuing WHO prequalification. Regulatory status must not be confused with a blanket endorsement of every future software configuration or every context in which it may be used.[1]

Keeping the clinical hand-off in the same visit

The wider programme has also provided more than 3,000 free screenings in under-resourced public clinics, according to Google, with a target of 10,000 sponsored screenings annually. Counts of scans make the expansion visible. The next important service measures are the proportion receiving confirmatory testing, the time to treatment and completion of treatment after a positive diagnosis.[1]

Google also frames the initiative as a move from a single tool toward adaptable foundation-model components. MedSigLIP aligns medical images with clinical text; MedGemma supplies open-weight multimodal capabilities that health systems can fine-tune and, in some configurations, run locally. Indonesia is described as preparing domestic adaptation of these models for a much larger screening ambition. Those national aspirations are not accomplished numbers, and they should remain separate from the specific South African deployment that is already operating.[1]

A screening service built around the setting

Offline operation is a practical advantage in a mine or clinic with unreliable connectivity. It reduces dependence on an external connection at the moment the image is taken. The programme still needs trained staff, maintained equipment and a dependable referral pathway, but the software can support the screening decision where the patient already is.[2][1]

At Mponeng, the innovation is the arrangement of care as much as the image classifier. Portable equipment, local processing and a route to confirmatory testing bring the stages closer together where miners work. Measuring how many flagged patients complete that pathway will show how effectively a faster screen becomes timely treatment.[1]

Report an error or suggest a correction