Google AI TB Screening Now Live in 40 Clinics
Google Research pairs its open medical models with a CE-certified X-ray service: roughly 40 facilities in six countries already screen with it.
In short
Google Research is putting its open medical models MedGemma, MedSigLIP and HeAR behind a CE-certified chest X-ray service from Nexus AI that the blog post says now runs in about 40 facilities across six countries.
At a glance
- Google puts TB's annual death toll at 1.23 million people.
- Nexus AI holds CE class IIb medical-device certification; WHO prequalification and FDA clearance are still pending.
- More than 90 clinicians have assessed over 25,000 chest radiographs.
- Turnaround falls from weeks to under 45 seconds, according to the post.
- MedGemma ships at 4B and 27B parameters and claims 10 million downloads.
Google Research used an October 23, 2026 post to lay out where its open medical models have landed: MedGemma, MedSigLIP and the audio model HeAR now sit behind a chest X-ray service from Pretoria-based Nexus AI, which the post says is running in roughly 40 facilities across six countries.
The post puts TB's annual death toll at 1.23 million. Taking the image is rarely the bottleneck; reading it is. The WHO has recommended computer-aided detection for TB screening since 2021.
The models behind the service
MedGemma ships in 4B and 27B parameter sizes, and Google claims 10 million downloads for the family. MedSigLIP is the lighter vision-language model for medical imaging, while HeAR, trained on more than 300 million audio segments per the post, listens to coughs instead of looking at films.
Open weights are load-bearing here, not a footnote. They let the models run on local hardware so scans never leave the clinic, which is the argument Google attaches to Indonesia's plan to screen 50 million citizens a year.
What the field numbers say
Nexus AI holds CE MDR class IIb certification under EU regulation 2017/745 plus ISO 13485 quality certification; WHO prequalification and U.S. FDA clearance are being pursued, not granted. More than 90 clinicians have assessed over 25,000 chest radiographs.
Turnaround drops from weeks to under 45 seconds, according to the post. Some 3,000 free screenings have gone to under-resourced public clinics, against a stated target of 10,000 sponsored screenings per year.
Where it is deployed
The service is live in South Africa, Sierra Leone and Côte d'Ivoire, with the Democratic Republic of the Congo, Zambia and Cameroon named as next. At the Mponeng mine in South Africa, the post describes ultra-portable X-ray kits and offline inference working four kilometers underground.
On money, Google lists a 250,000 dollar Google.org grant to the Aurum Institute and 1.5 million dollars in regional collaborative funding for HELINA. Partners include the Stop TB Partnership, FIND, UNOPS and South Africa's Department of Health.
Open questions
The accuracy evidence predates the rollout: a 2019 benchmark in Radiology and a September 2024 validation in NEJM AI reporting high sensitivity across diverse patient groups. The post gives no sensitivity or specificity figures for the live deployment.
Every figure above comes from Google's own blog, and we had no independent source to confirm it. How many of the more than 25,000 reads produced a confirmed diagnosis is also unstated.
FAQ
Which AI models does Google use for TB screening?
Per the blog post, MedGemma at 4B and 27B parameters, the vision-language model MedSigLIP, and HeAR, a cough model trained on more than 300 million audio segments. The field X-ray reading itself is delivered by partner Nexus AI.
Is the TB software approved as a medical device?
Nexus AI carries CE class IIb certification under EU regulation 2017/745 and ISO 13485 quality certification. WHO prequalification and U.S. FDA clearance are being pursued but were not in hand at the time of the post.
Which countries are screening with it today?
South Africa, Sierra Leone and Côte d'Ivoire are live, roughly 40 facilities across six countries in total. Google names the Democratic Republic of the Congo, Zambia and Cameroon as next, and Indonesia plans 50 million screenings a year.
