Tuberculosis kills more than 1.23 million people a year, and the bottleneck in many high-burden places is not treatment but finding the cases: there are too few radiologists to read chest X-rays. Google Research has set out, at the IDWeek 2026 conference, how its TB work has moved from research papers to clinics, and how it now rests on open models that local health systems can run themselves.
The clinical route runs through Nexus AI, a medical technology company in Pretoria that turned Google's chest X-ray algorithms into a certified product. It holds Class IIb certification under the EU Medical Device Regulation, is pursuing World Health Organization prequalification through an agreement with FIND, and is on a path towards US FDA clearance. The tool is in use in South Africa, Sierra Leone and Côte d'Ivoire, with expansion under way in the Democratic Republic of Congo, Zambia and Cameroon. Across about 40 facilities, more than 90 clinicians have reviewed over 25,000 X-rays, with results in under 45 seconds instead of weeks. One deployment pairs a portable X-ray unit with the AI at Mponeng, one of the world's deepest gold mines.
The open part is what makes it scale. Google has released MedSigLIP, which links medical images to the language used to describe them, and MedGemma, open-weight models at 4B and 27B parameters that we review in its 4B form. Indonesia, which Google says has the world's second-highest TB burden and wants to screen 50 million people a year, is fine-tuning both on its own X-rays so patient data stays in the country. For places without imaging, HeAR, trained on more than 300 million audio segments, screens from cough sounds recorded on a phone.
For anyone weighing medical AI, this is a useful case of a model that runs offline and can be adapted locally rather than sent to a foreign cloud. MedGemma has passed 10 million downloads, by Google's count.
What the post does not give is the number that matters most: how accurate the deployed tools are. There are no sensitivity or specificity figures for the X-ray product or for HeAR, which is still building clinical evidence with partners. WHO prequalification and FDA clearance are pending, and the throughput figures come from Google and its partner. Until independent accuracy data appears, the scale of deployment is clearer than its clinical effect.