Access to timely healthcare often depends on geography: where people live, what local clinics can provide, and whether specialists are available. This disparity is especially stark in the fight against tuberculosis (TB). Although TB is preventable and curable, it still causes more than 1.23 million deaths per year. A major barrier is radiology: in high-burden settings—from remote rural communities to deep underground mines—radiologists and imaging infrastructure are scarce, leaving millions of active cases undetected.
Demonstrating AI can read chest radiographs for TB screening
Teams at Google Research and Google for Health asked whether AI tools can interpret chest radiographs (CXRs) for TB screening with expert-level accuracy. Chest radiography is the frontline imaging modality for pulmonary screening, but detecting subtle TB signs—such as apical infiltrates and cavitary lesions—requires specialty expertise rarely available in local clinics. In 2019 the group published a benchmark in Radiology showing deep learning algorithms can detect active pulmonary TB at performance levels comparable to experienced radiologists.
Independent verification is needed for real-world use. In August 2021, Stop TB Partnership evaluated automated computer-aided detection (CAD) software across diverse patient cohorts in The Lancet Digital Health and ranked the team’s model among the top-performing TB screening algorithms globally. That year WHO recommended CAD for TB screening and triage. In September 2024, the group published a prospective, multi-site validation in NEJM AI showing AI-assisted screening maintains high sensitivity across diverse populations while also reliably identifying incidental pulmonary and cardiac abnormalities.
Bridging the translation gap from model to clinical tool
Publishing research is only the first step. Between an algorithm and a trusted bedside tool lies a "translation gap": ensuring software works across different X-ray machines and offline settings, integrating it into clinical workflows, and meeting medical device regulations.
To address this, Google partnered with Nexus AI, a Pretoria-based medical technology company, which transformed the chest X-ray algorithms into a certified clinical diagnostic product. Regulatory navigation involved extensive audits: Nexus AI achieved CE MDR Class IIb certification under the EU Medical Device Regulation (EU MDR 2017/745) and obtained ISO 13485 certification for medical device quality management. Nexus AI also signed an agreement with FIND to pursue World Health Organization Pre-qualification (WHO PQ) and is on a pathway toward U.S. FDA clearance.
Frontline impact: delivering screening where patients are
Nexus Intelligence is deploying the technology in locations with the greatest diagnostic need. Across six countries—South Africa, Sierra Leone, and Côte d'Ivoire, with expansion planned in the Democratic Republic of the Congo, Zambia, and Cameroon—the tool is active in roughly 40 healthcare facilities. More than 90 clinicians have evaluated over 25,000 chest radiographs to date; turnaround times that used to take weeks are now under 45 seconds, enabling rapid triage while clinicians make final decisions.
The collaboration includes an equity commitment: Nexus AI has provided over 3,000 free screenings to under-resourced public clinics, aiming for 10,000 sponsored screenings per year.
Screening at the extremes: Mponeng Gold Mine deployment
A notable deployment is at Mponeng Gold Mine in South Africa, one of the world’s deepest mines. Deep-shaft miners face chronic silica exposure and elevated risk of pulmonary disease. In partnership with Stop TB Partnership, Harmony Gold Mine, and the South African Department of Health, Nexus AI deployed ultra-portable digital X-ray kits and offline AI four kilometers underground. Miners are screened on-site within seconds; those flagged with presumptive TB are immediately referred for confirmatory molecular testing. The combined screening, triage and confirmatory testing can now occur in a single visit instead of over weeks.
From point solutions to open foundations: HAI-DEF and MedGemma
Single-partner, bespoke solutions cannot alone address global TB burden. To scale, the team shifted toward open, adaptable foundations so developers and health ministries can build locally tailored solutions.
Under the Health AI Developer Foundations (HAI-DEF), the group open-sourced chest X-ray embeddings and MedSigLIP, a lightweight vision-language model trained to align medical images (such as chest X-rays) with clinical language. Because MedSigLIP understands both images and text, developers can classify or search images with plain-language labels or fine-tune it for specific tasks using only a small set of local examples. This lets local teams reach strong performance on modest regional datasets without massive compute.
Building on that, they released MedGemma, a family of open-weight multimodal models (4B and 27B parameter configurations). MedGemma goes beyond basic classification to multimodal clinical reasoning: it synthesizes radiographic findings with clinical history, answers clinician queries in natural language, and runs locally on-device without internet access. Since launch, MedGemma has been downloaded more than 10 million times worldwide.
Sovereign scale: supporting Indonesia’s national TB initiative
Open models help public health systems build sovereign solutions that keep patient data local. Indonesia, which carries the world’s second-highest TB burden, set an ambitious target: screening 50 million citizens annually to eliminate the disease. Reviewing 50 million X-rays manually is infeasible, and proprietary cloud software raises data residency concerns. The Indonesian Ministry of Health is fine-tuning MedGemma and MedSigLIP on domestic chest radiographs so models run on local infrastructure and patient data remain within national borders, giving health leaders control over clinical deployment.
Complementary screening: bioacoustics with HeAR
Because imaging cannot reach every remote health post, Google Research developed Health Acoustic Representations (HeAR). Trained on over 300 million audio segments, HeAR enables non-invasive acoustic screening from cough sounds recorded on a standard smartphone microphone.
HeAR is being used in collaborative research across sub-Saharan Africa with CIDRZ (Zambia), the Aurum Institute (Ghana, supported by a $250,000 Google.org grant), and HELINA’s regional collaboration (supported with $1.5 million). Combined, acoustic triage (HeAR), mobile X-ray screening (HAI-DEF), and multimodal clinical reasoning (MedGemma) provide an integrated diagnostic cascade for underserved communities.
Looking ahead
Moving from a lab algorithm to a CE-certified tool in clinics across six countries has required rigorous science, open developer platforms, and strong local partnerships. Partners familiar with their communities are expanding these efforts: Nexus AI is entering new countries, Indonesia is integrating TB screening into its national program, and CIDRZ, the Aurum Institute and HELINA are building clinical evidence for cough-based triage. The researchers’ role is to support these partners with open models, foundational research and ongoing assistance so more people with TB are diagnosed and treated earlier.
Acknowledgements
The authors thank collaborators at Nexus AI, Stop TB Partnership, UNOPS, CIDRZ, Aurum Institute, HELINA, Harmony Gold Mine, the South African Department of Health, and the Indonesian Ministry of Health for their clinical leadership and commitment to global health equity.
