TB screening stops waiting for a radiologist: across about 40 sites in six countries, reading time falls from weeks to under 45 seconds

The bottleneck in TB screening has moved from algorithm accuracy to whether a tool runs reliably in real clinics: a chest X-ray reading tool is now in routine use at about 40 healthcare facilities across six countries including South Africa. Such tools previously stayed at the research-benchmark and paper stage, while the distance from an algorithmic model to a trusted bedside tool includes adapting to varied X-ray machines, offline operation, clinical workflow integration and medical device regulation. Nexus AI, a Pretoria-based medical technology company, translated the chest X-ray algorithms into a CE-certified clinical diagnostic tool; more than 90 clinicians have since evaluated over 25,000 chest radiographs with it.

The change in reading speed lets triage finish within a single visit instead of losing patients during weeks of waiting. In high-burden settings, imaging infrastructure was often days away and radiologists scarce, so millions of active cases went undetected. Across about 40 facilities in six countries, turnaround fell from weeks to under 45 seconds; at the Mponeng Gold Mine in South Africa, ultra-portable digital X-ray kits and offline AI were deployed four kilometers underground, screening miners on-site within seconds and referring flagged cases immediately for confirmatory molecular testing.

The same capability is handed to health systems as open models, so local teams can build their own screening solutions from modest regional data. A custom solution with a single partner cannot address the global TB burden alone, so the route shifted from isolated deployments to open, adaptable foundations. The Health AI Developer Foundations released chest X-ray embeddings and MedSigLIP, a lightweight vision-language model that aligns medical images with the clinical language describing them; MedGemma, built on this, offers open-weight multimodal configurations at 4B and 27B parameters, runs locally on-device without internet, and has seen more than 10 million downloads worldwide since launch.

Indonesia's Ministry of Health is fine-tuning MedGemma and MedSigLIP on domestic chest radiographs to bring TB screening into its national program, with a goal of screening 50 million citizens annually. Reviewing 50 million X-rays manually is infeasible, and proprietary cloud software raises data residency concerns. Because the models run on local infrastructure, patient data stays within national borders; separately, HeAR, trained on over 300 million audio segments, explores non-invasive acoustic screening from cough sounds recorded on a standard smartphone microphone, with collaborative research under way in sub-Saharan Africa.

The report covers the translation path Google Research and Google for Health shared at IDWeek 2026: from the 2019 Radiology benchmark, the 2021 independent Stop TB Partnership evaluation and the 2024 NEJM AI prospective multi-site validation, to Nexus AI's CE-certified tool deployed at about 40 facilities in six countries, plus open models including HAI-DEF, MedSigLIP, MedGemma and HeAR. The material states that Nexus AI is expanding to new countries, Indonesia is integrating screening into its national program, and CIDRZ, the Aurum Institute and HELINA are building clinical evidence for cough-based triage.

The material positions AI as triage support, with final clinical decisions left to health workers, and says AI-assisted screening maintains high sensitivity while also detecting other incidental pulmonary and cardiac abnormalities, but it reports no sensitivity or specificity values for this six-country deployment. Cough-based acoustic screening remains collaborative research with clinical evidence still accumulating; Indonesia's 50-million-a-year screening figure is its Ministry of Health's goal, not a completed screening count.