Google Partner's AI Has Read 25,000 TB X-Rays

Four kilometres underground at South Africa’s Mponeng gold mine, miners are having their chests X-rayed and the images read by AI on the spot. It is one of about 40 health facilities where a screening tool built on Google’s chest X-ray model is now used, Google Research said on Thursday 23 October, with more than 25,000 X-rays read so far.

The update, shared at the IDWeek infectious diseases conference in Washington, DC, sets out how Google has moved from research papers to deployed tools and open models for tuberculosis, a disease that killed 1.23 million people in 2024, according to the World Health Organization. The figures are Google’s own. The tool flags people for follow-up testing; it does not make the diagnosis, and clinicians make the final decision.

In brief

What’s new: Google says a Pretoria company’s AI tool, built on its model, has read 25,000+ chest X-rays in about 40 facilities, and Indonesia is tuning Google’s open MedGemma models for national TB screening.
Why it matters: Radiologists are scarce where TB is most common, and AI triage can cut waits for X-ray results from weeks to seconds.
What’s unknown: How many extra cases the tool has found, and how accurate it has been in the field.

A Pretoria company turned Google’s research into a certified device

Google’s work began with a benchmark published in Radiology in 2019, showing deep learning could spot active pulmonary TB on chest X-rays at the level of experienced radiologists. In 2021 an evaluation by the Stop TB Partnership in The Lancet Digital Health ranked the model among the top-performing TB screening algorithms, and in 2024 a prospective study in NEJM AI found it kept high sensitivity across varied patient groups, Google says.

Turning that into a clinical product fell to Nexus AI, a medical technology company in Pretoria. Its tool has CE Class IIb certification under the EU’s Medical Device Regulation and is certified to the ISO 13485 quality standard. Nexus has also signed an agreement with the diagnostics nonprofit FIND to pursue WHO prequalification, which international buyers rely on, and is working towards US FDA clearance.

The tool now runs in about 40 facilities across South Africa, Sierra Leone and Côte d’Ivoire, with expansions under way in the Democratic Republic of the Congo, Zambia and Cameroon. More than 90 clinicians have used it to evaluate over 25,000 chest X-rays, and Google says it cuts the wait for a result “from weeks to under 45 seconds”. Nexus has given more than 3,000 free screenings to under-resourced public clinics and aims for 10,000 sponsored screenings a year.

The 25,000 figure is not new. A Google blog in June gave the same total across 40 locations in six countries.

Miners can now be screened, triaged and tested in one visit

Mponeng is one of the deepest mines in the world, and its miners face long-term exposure to silica dust and a higher risk of lung disease. Working with the Stop TB Partnership, Harmony Gold and South Africa’s Department of Health, Nexus took ultra-portable digital X-ray kits and AI that runs offline down the shaft. Miners flagged with presumptive TB can be sent straight for confirmatory molecular testing, so “screening, triage and confirmatory testing can now happen in a single visit”, compressing a process that often takes weeks, according to Google.

This is the same logic as other point-of-care AI screening TSN has covered, such as the Johns Hopkins trial in which AI eye exams at routine diabetes visits lifted screening rates: bring the test to the patient and fewer drop out.

Indonesia wants to screen 50 million people a year, keeping data at home

Google’s bigger bet is on open models that health systems can adapt themselves. It has released MedSigLIP, a lightweight model that links medical images to clinical language, and MedGemma, a family of open-weight models in 4-billion and 27-billion parameter sizes that can run on local hardware without an internet connection. Google says MedGemma has been downloaded more than 10 million times.

Indonesia is the clearest test. It has the world’s second-highest TB burden, accounting for 10% of new cases in 2024 according to the WHO, and its Ministry of Health aims to screen 50 million people a year. The ministry is fine-tuning MedGemma and MedSigLIP on Indonesian chest X-rays and running them on local infrastructure, so patient data stays inside the country. Indonesian developers have already built on Google’s open models elsewhere, as with GoTo’s 27-billion-parameter Gemma-based model for local languages.

Google is also backing cough-based screening for places with no X-ray machine. Its HeAR model, trained on more than 300 million audio clips, is being studied by CIDRZ in Zambia, the Aurum Institute in Ghana with a $250,000 Google.org grant, and a $1.5 million regional collaboration led by HELINA.

The deployment is real, but the evidence still has to catch up

The strongest part of Google’s account is the chain it describes: peer-reviewed research, an independent ranking, a certified device and a local company that sells and supports it. Too many health AI projects stop after the first paper. Deployment in the mine, offline, shows the tool was built for the places where radiologists are scarcest, and keeping clinicians in charge of the decision is the right design for a triage tool.

What the post lacks is outcome data. It reports X-rays read and seconds saved, but not how many TB cases were found that would otherwise have been missed, how many flagged patients were confirmed, or how the tool performed in the field against its published benchmarks. Those numbers matter more than throughput, and the unchanged 25,000 total since June suggests either slow growth or a figure that has not been updated. That is similar to the caution TSN noted with Microsoft’s diagnostic AI, where the step to testing with real patients is still to come.

Indonesia is the case to watch. Screening 50 million people a year with a nationally tuned open model, on national infrastructure, would be a large test of whether open models can serve public health at scale. The next markers are the ministry’s own results and whether Nexus wins WHO prequalification, which would open the door to international procurement.