Medical Technology · global
Bringing Tuberculosis Screening to Mines and Remote Communities: Google Reports Progress in Chest X-ray AI Deployment
From underground mines in South Africa to local model tuning in Indonesia, chest X-ray AI is entering tuberculosis screening settings. But how quickly images are analyzed, how widely systems are deployed, and whether they help more patients receive an early diagnosis remain separate questions.
In places with a shortage of radiologists, obtaining a chest X-ray is only the starting point for tuberculosis screening. Whether images can be interpreted promptly and people with suspected tuberculosis can access confirmatory testing determines whether screening truly works. Google Research released an update on October 23 to coincide with the IDWeek conference, outlining partners’ experiences bringing imaging AI into clinics, mobile screening programs, and mines, while also introducing Indonesia’s approach to using open medical models.
South African company Nexus Intelligence licensed Google’s chest X-ray model to develop Nexus AI CXR. The system runs two models in parallel: one highlights areas that deviate from the appearance of normal lungs, while the other looks for imaging features suggestive of pulmonary tuberculosis. Heatmaps and contour lines then help healthcare professionals interpret the images. The company says each image takes less than 45 seconds to analyze and that the system can operate offline; this is the speed of assisted triage and should not be treated as the time needed to confirm a tuberculosis diagnosis.
In this update, Google said the system had evaluated more than 25,000 chest X-rays at approximately 40 healthcare sites. However, the same image count and number of sites had already appeared in a Google article on June 16, which also said the deployment covered six countries. These figures therefore describe the existing deployment scale and cannot be interpreted as new screening results from October or directly converted into a count of confirmed cases.
South Africa’s Mponeng gold mine offers a concrete example. According to an account Nexus published in April, the team worked with Stop TB Partnership, Harmony Gold, and South Africa’s Department of Health on March 26 to bring MinXray ultraportable X-ray equipment and AI to nearly four kilometers underground, allowing miners to undergo screening at their workplace. Google noted that people with suspected tuberculosis could be referred for molecular testing; the value of this type of deployment lies in shortening the distance between detecting an abnormality and arranging follow-up testing.
Clinical evidence still needs to be considered separately from deployment accounts. Google reviewed existing independent algorithm evaluations and a prospective multicenter validation published in 2024, but this update did not report sensitivity, specificity, the proportion of confirmed cases, or referral completion for these field screenings. The number of images processed can indicate the scale of use, but it is insufficient to demonstrate that missed diagnoses have decreased or that patients are starting treatment earlier.
Certification also has clear boundaries. Google and Nexus both said Nexus AI CXR had obtained Class IIb CE MDR certification under the European Union’s Medical Device Regulation, and the company also lists ISO 13485:2016 quality management certification. However, Nexus lists the addition of lesion description labels and a feature that uses a large language model to generate reports as beta features, and says it is applying to expand the scope of its CE certification. Certification of the existing screening product cannot be interpreted as approval of all new features.
Another approach is taking shape in Indonesia. According to Google, Indonesia’s Ministry of Health is tuning MedGemma and MedSigLIP using domestic chest X-ray data so the models can run on local infrastructure, addressing the need to keep patient data within the country. This is at a different stage from the clinical deployment of Nexus’s existing product; the update provided no validation results for the versions tuned in Indonesia. Moving from available models to reliable public health services still requires confirming their performance in local populations and on local equipment, and ensuring that people flagged as suspected cases actually receive testing and care.