将基础模型带到结核病筛查前线

Preeti Singh, Lead, Technical Program Manager, and Shravya Shetty, Distinguished Software Engineer, Google Research Tuberculosis remains one of the world's deadliest infectious diseases, claiming an estimated 1.23 million lives each year. Despite decades of global investment, too many cases go undetected in the communities that carry the highest burden. Over the past five years, we’ve worked to change that trajectory, moving from early research benchmarks toward open foundation models that could be tuned and deployed at the point of care. Getting access to timely healthcare often comes down to an accident of geography: where you live, what resources your local clinic possesses, and whether a specialist is within reach. Nowhere is this disparity more acute than in the fight against tuberculosis (TB). Despite being preventable and curable, TB claims more than 1.23 million lives annually. A primary bottleneck is radiology: For those in high-burden settings, from people who live in remote rural areas to miners working deep underground, radiologists are scarce, and imaging infrastructure is often days away. As a result, millions of active cases go undetected, perpetuating transmission. Over the past several years, our team across Google Research and Google for Health has worked to change that trajectory: moving from an AI-based tool that builds on our existing work in medical imaging to identify people with presumptive TB for follow-up testing to prospective clinical trials, regulatory clearance with local partners, and ultimately open foundation models like MedGemma that empower public health systems worldwide. This week at IDWeek 2026 in Washington, D.C., we are sharing how we’re bringing health AI research to reality, and how foundational models can democratize TB screening at global scale. Our work began with a fundamental question: Can AI-based tools interpret chest radiographs (CXRs) with the accuracy of an expert radiologist for TB screening? Chest radiography is the frontline imaging modality for pulmonary screening, but detecting subtle signs of TB like apical infiltrates (fluid in the top portion of the lung) and cavitary lesions (tissue damage) requires specialized expertise that is rarely available in rural clinics. In 2019, we published a benchmark in Radiology , demonstrating that deep learning algorithms could detect active pulmonary TB at performance levels matching experienced radiologists. Real-world utility, however, demands independent verification. In August 2021, the Stop TB Partnership evaluated automated computer-aided detection (CAD) software across diverse patient cohorts in The Lancet Digital Health , ranking our model among the top-performing TB screening algorithms globally. That year, WHO also recommended CAD for TB screening and triage. In September 2024, in NEJM AI , we detailed a prospective, multi-site validation demonstrating that AI-assisted screening maintains high sensitivity (minimizing missed cases) across diverse patient populations, while also reliably detecting other incidental pulmonary and cardiac abnormalities. In medical AI, publishing a paper is only the beginning. Between an algorithmic model and a trusted bedside tool lies the "translation gap": making software reliable across varied X-ray machines and offline settings, integrating it into clinical workflows, and meeting medical device regulations. With this in mind, we partnered with Nexus AI , a Pretoria-based medical technology company, and they translated our chest X-ray algorithms into a certified clinical diagnostic tool. Navigating the regulatory landscape required extensive audits: CE MDR class IIb certification : Under the European Union’s Medical Device Regulation ( EU MDR 2017/745 ), Nexus AI achieved Class IIb certification — a rigorous designation that verifies the tool is safe for real-world patient care based on prospective clinical evaluations. Quality standards : Nexus AI obtained certified compliance wit