The Pulse
Google Brings Open Medical Models to TB Screening
Google says its chest X-ray screening tool is operating in about 40 health facilities across six countries, including an offline deployment four kilometers underground at South Africa’s Mponeng Gold Mine. The company also describes open mod
AI.info Team · Written with AI · fact-checked ·
“Screening, triage and confirmatory testing can now happen in a single visit, compressing a pathway that often takes weeks.”
Preeti Singh, Lead, Technical Program Manager, and Shravya Shetty, Distinguished Software Engineer, Google Research
Google Research says its tuberculosis screening work has moved from research models to a certified chest X-ray tool deployed in clinics, while newer open models are being adapted for local health systems. In a post published October 23, the company described deployments with Pretoria-based Nexus AI and separate efforts to tune MedGemma and MedSigLIP for national screening needs.
Nexus AI’s tool reaches about 40 facilities
Nexus AI’s chest X-ray tool operates in approximately 40 healthcare facilities across six countries, Google says. Deployments are underway in South Africa, Sierra Leone and Côte d’Ivoire, with expansion underway in the Democratic Republic of the Congo, Zambia and Cameroon. More than 90 clinicians have assessed over 25,000 X-rays, and Google says screening turnaround has fallen from weeks to under 45 seconds.
The system flags findings that may indicate tuberculosis or require further investigation; clinicians retain responsibility for decisions. Google says Nexus AI has delivered more than 3,000 free screenings at under-resourced public clinics and is working toward 10,000 sponsored screenings each year. The company describes the tool as a triage aid, not a replacement for confirmatory testing.
Offline screening goes four kilometers underground
At South Africa’s Mponeng Gold Mine, Nexus AI has deployed portable digital X-ray kits and offline AI four kilometers underground, working with the Stop TB Partnership, Harmony Gold Mine and the South African Department of Health. Miners can be screened on-site, and people flagged as having presumptive TB can be referred for confirmatory molecular testing. Google says screening, triage and confirmation can take place in one visit.
The setting illustrates why offline operation matters to the deployment: workers are screened at the mine rather than having to rely on imaging infrastructure elsewhere. Google also notes miners’ silica dust exposure and elevated pulmonary disease risk. The post does not report how many people have been screened at Mponeng or provide site-specific accuracy results.
Open models are a separate route to local adaptation
Google says it has released chest X-ray embeddings and MedSigLIP through its Health AI Developer Foundations program. MedSigLIP aligns medical images with clinical language, allowing developers to classify or search images with plain-language labels and fine-tune for specific tasks using a small set of local examples. Google presents these as building blocks for teams adapting models to regional data, not as evidence that every local adaptation is already clinically validated.
The company also describes MedGemma, an open-weight family with 4-billion- and 27-billion-parameter configurations. Google says the models can combine radiographic findings with clinical history and run locally without an internet connection. That is a different step from Nexus AI’s certified screening system: the post does not say MedGemma itself is the model running in those 40 facilities.
Indonesia plans to screen 50 million people a year
Indonesia’s Ministry of Health is fine-tuning MedGemma and MedSigLIP on domestic chest X-rays as part of a goal to screen 50 million citizens annually, according to Google. Running the models on local infrastructure is intended to keep patient data within the country and give health officials control over deployment. The post describes the work as underway; it does not report completed national screening results.
Google’s account therefore combines a deployed clinical tool with open models still being adapted for broader use. It gives concrete operational figures for Nexus AI’s reach and speed, but no comparable performance figures for Indonesia’s fine-tuned models. The distinction matters: access to downloadable models can support local development, while evidence of safe and effective screening still depends on validation in the settings where those systems will be used.