ModelsArchitectures & capability
Google details its TB screening program: Nexus AI's X-ray tool reaches about 40 sites, and Indonesia adapts open MedGemma models
Google's TB screening work has moved from published studies to a certified product: Nexus AI's X-ray tool, built on a licensed Google model, runs at about 40 sites in six countries, while Indonesia is adapting Google's open MedSigLIP and MedGemma models.
From the Google Research article of October 23, 2026 and cited sources: The clinical evidence comes from peer-reviewed studies. [2] [3]
SOURCES[2] Radiology: Kazemzadeh et al. pubs.rsna.org[3] NEJM AI study via PMC: Prospective Multi-Site Validation pmc.ncbi.nlm.nih.gov
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Google Research article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2507.05201 (Jul 2025) [4] date from arXiv ID
- When
- October 23, 2026 article published
- Covered
- Oct 24 edition
- Evidence
- 15 sources across 11 sites
- 1Bringing foundational models to the front lines of TB…research.google
- 2Radiology: Kazemzadeh et al.pubs.rsna.org
- 3NEJM AI study via PMC: Prospective Multi-Site Validationpmc.ncbi.nlm.nih.gov
- 4MedGemma Technical Report (arXiv)arxiv.org
- 5Daily News: Pretoria-developed AI finding hidden TBdailynews.co.za
- 6Google Africa blog: Meet the Pretoria teamblog.google
Reach: not measured yet
The clinical evidence comes from peer-reviewed studies. A 2022 Radiology study using data from 10 countries found higher sensitivity than radiologists, with specificity that was not significantly worse. A three-site prospective study in Zambia reported 87% sensitivity and 70% specificity at the high-sensitivity threshold. At the threshold set to resemble radiologists, it reported 78% and 82%, against the radiologists' 76% and 82%. [2] [3]
Nexus AI, a Pretoria startup, turned the licensed model into a CE Class IIb device. It pairs an abnormal-lung model with a TB-specific model, runs offline, and returns results with heatmaps in about 45 seconds. A clinician makes the final decision. The deployment figures are vendor-reported: more than 25,000 X-rays read at roughly 40 sites, including one in an underground gold mine. WHO prequalification is still pending. [1] [6]
Google is also encouraging countries to build their own TB tools on its open models. It cites Indonesia, which is fine-tuning the models on local X-rays with the aim of screening 50 million people a year. MedSigLIP works at 448×448 pixels and averages a zero-shot AUC of 0.844 on chest X-ray findings, with results as low as 0.650 on some findings. No performance results for the Indonesian fine-tunes have been published. [1] [4] [8]
In the Zambia prospective study, the model's performance was close to radiologists'.
Read the full assessment
If open models can be adapted locally, health ministries could build their own screening tools instead of buying them, but validation and regulatory work would then fall on local teams.
Sources are gathered, including a few that conflict with the post. Writing the dossier now.
Executive brief
Google's TB work has gone from research papers to a CE-certified device. Google says Nexus AI, a Pretoria startup that licensed Google's chest X-ray model, has read more than 25,000 X-rays at about 40 sites in six countries. One of those sites is 4 km underground in a gold mine. Google now wants others to build their own TB tools on its open models, MedSigLIP and MedGemma. Its main example is Indonesia, which Google says is fine-tuning these models on its own X-rays to screen 50 million people a year. No source outside Google confirms the Indonesia project. Two research milestones in Google's own account don't match the published record.
What changed and event timeline
Google TB model matches radiologists in Radiology
Kazemzadeh et al. tested a deep-learning system on data from 10 countries. It had higher sensitivity than nine radiologists (88% vs 75%) and specificity that was not significantly worse ().
Prospective study in Zambia
NEJM AI reported on 1,910 analyzed patients at three sites. The TB AI reached 87% sensitivity and 70% specificity at its high-sensitivity setting ().
MedGemma and MedSigLIP released
A technical report described medical vision-language models built on Gemma and SigLIP, with chest X-ray benchmark results ().
South African press covers Nexus AI
Local reporting describes the Pretoria-built tool finding hidden TB cases ().
Google sets out the full TB program at IDWeek
The post covers the Nexus deployments, the Indonesia project, more than 10 million MedGemma downloads, and cough-based screening ().
Capabilities and access
- Nexus AI CXR: a CE MDR Class IIb device built on Google's licensed model. The clinician makes the final call (Google Africa blog).
- Open models (Google's HAI-DEF program): chest X-ray embeddings, MedSigLIP, and MedGemma in 4B and 27B sizes (MedSigLIP model card).
- HeAR: an audio model trained on about 300 million clips, including roughly 100 million coughs (Google blog).
Read the full section
- Nexus AI CXR: a CE MDR Class IIb device built on Google's licensed model. It runs two models side by side: one flags abnormal lungs, the other flags findings suggestive of TB. It works offline, shows heatmaps, and gives results in about 45 seconds. The clinician makes the final call (Google Africa blog). WHO prequalification and FDA clearance are still in progress.
- Open models (Google's HAI-DEF program): chest X-ray embeddings, MedSigLIP, and MedGemma in 4B and 27B sizes (MedSigLIP model card).
- HeAR: an audio model trained on about 300 million clips, including roughly 100 million coughs (Google blog).
Technical analysis for researchers and developers
- How the 2022 model was trained: pretraining on large chest X-ray sets, attention pooling, and noisy-student semi-supervised learning.
- Prospective study design: model scores were turned into yes/no using two thresholds, one tuned for high sensitivity and one set to resemble radiologists.
- MedSigLIP: works at 448×448 pixels, against 1280×1280 for Google's older CXR Foundation model.
Read the full section
- How the 2022 model was trained: pretraining on large chest X-ray sets, attention pooling, and noisy-student semi-supervised learning. On a four-country test set it scored AUC 0.89 (Radiology).
- Prospective study design: model scores were turned into yes/no using two thresholds, one tuned for high sensitivity and one set to resemble radiologists. Ten radiologists read the same images without knowing the true diagnosis (PMC).
- MedSigLIP: works at 448×448 pixels, against 1280×1280 for Google's older CXR Foundation model. Even so, its zero-shot AUC on chest X-ray findings averaged 0.844 (range 0.650–0.931) (arXiv).
- Running locally: in-country fine-tuning and offline inference are what make keeping data inside national borders possible. No sources report how well the Indonesian fine-tunes perform.
Claims and evidence
- Prospective results (independently checkable): sensitivity/specificity of 87%/70% and 78%/82% at the two thresholds, against radiologists' 76%/82%.
- The "Radiology 2019" claim conflicts with the record. Google's TB paper in Radiology appeared in 2022 (Google Research pub).
- The 2021 Stop TB ranking is not corroborated.
Read the full section
- Prospective results (independently checkable): sensitivity/specificity of 87%/70% and 78%/82% at the two thresholds, against radiologists' 76%/82%. Note that Google and CIDRZ, its Zambian partner, wrote the paper (PMC).
- The "Radiology 2019" claim conflicts with the record. Google's TB paper in Radiology appeared in 2022 (Google Research pub). The well-known 2019 TB paper in Radiology is from a different group (PMC6695514).
- The 2021 Stop TB ranking is not corroborated. Coverage of the Lancet Digital Health study lists five products: CAD4TB, InferRead, JF CXR-1, Lunit and qXR (TB Online). Google isn't among them, and Google isn't named among the top performers in the 12-product study in Scientific Reports either (PMC).
- Deployment figures, the Indonesia project and the 10 million downloads come only from Google, and the Nexus figures are repeated in Google's own Africa blog.
Context and prior work
WHO estimates 1.23 million TB deaths in 2024. That is the lowest ever recorded, but well short of WHO's targets (WHO). The X-ray reading software market is crowded. Qure.ai, Lunit and Delft performed best in the independent 12-product study (PMC), and Indonesia has already trialled Lunit's tool (Lunit). For cough screening, India's Salcit already uses HeAR in its Swaasa app (Inquirer).
Limitations, safety and contested findings
- At its high-sensitivity setting, the TB model's specificity was 70%. That means many false positives, each needing confirmatory testing (PMC).
- The prospective evidence comes from three sites in Zambia, a high-HIV setting. Results may differ elsewhere.
- MedSigLIP's zero-shot AUC drops as low as 0.650 on some findings (arXiv).
Read the full section
- At its high-sensitivity setting, the TB model's specificity was 70%. That means many false positives, each needing confirmatory testing (PMC).
- The prospective evidence comes from three sites in Zambia, a high-HIV setting. Results may differ elsewhere.
- MedSigLIP's zero-shot AUC drops as low as 0.650 on some findings (arXiv).
- The 25,000 X-rays count how many images were read. No outcome data from the deployment, such as cases found or confirmed, has been published.
- The two research-timeline discrepancies above are unresolved.
Business and practitioner implications
- Getting to market: a licensed research model only became a product through local certification (CE Class IIb, ISO 13485).
- Competition: open-weight models let health ministries build their own tools instead of buying from established vendors.
- For builders: check MedSigLIP's chest X-ray performance on your own population before deploying.
Read the full section
- Getting to market: a licensed research model only became a product through local certification (CE Class IIb, ISO 13485). Large-scale purchasing through global agencies depends on WHO prequalification, which Nexus is still pursuing.
- Competition: open-weight models let health ministries build their own tools instead of buying from established vendors. This undercuts commercial CAD sellers, but it also hands validation and regulatory work to whoever does the fine-tuning.
- For builders: check MedSigLIP's chest X-ray performance on your own population before deploying. The reviewed sources show no independent evaluation of how the open models perform on TB.
The source trail.
Sources (15)
Bringing foundational models to the front lines of TB screening
Article text retrieved; extracted text may omit tables or interactive elements.
research.google