A recent update brings together different kinds of tools

Google Research’s October 23 account of tuberculosis work connects chest-X-ray screening with local deployments and open health AI models. The company describes Nexus AI’s partner-led screening work and the use of adaptable models such as MedGemma and MedSigLIP.

These are not interchangeable achievements. An application used in a screening service, a model available to developers and a person receiving a confirmed diagnosis each represent a different stage.

The update is a useful prompt to separate those stages. It should not be read as a new finding that any downloaded health model can diagnose tuberculosis.

Screening identifies who needs further investigation

Computer-aided detection, or CAD, can help interpret a chest X-ray during TB screening. The World Health Organization’s June 2025 update reported that six assessed CAD products met its performance standards for screening people aged 15 and older. It did not extend that recommendation to younger children.

WHO also states that people with a positive screening result should undergo confirmatory testing before starting TB treatment. A screening signal is therefore not the end of the diagnostic pathway.

That age limit and testing step matter when reading broad claims about access. Evidence for a specified product and population does not automatically establish the same performance for another model or another group.

A foundation model is a component, not a finished service

Google’s MedGemma documentation describes a starting point for developers building healthcare applications. Its model card says developers are responsible for adapting the model to their intended use and validating the resulting application.

The limitations are explicit: its generated outputs are not intended to directly inform clinical diagnosis, patient-management decisions or treatment recommendations. Outputs are preliminary and need to be checked independently against clinical information.

The card also asks developers to validate with data representative of the intended setting, including patient characteristics and imaging devices. A strong baseline on a benchmark cannot answer all of those local questions.

Judge the whole pathway, not just the model’s response

A fast image analysis can shorten one task without showing how long it takes a person to reach confirmatory testing and care. Likewise, running a model locally addresses where computation happens; it does not by itself establish clinical reliability.

For readers assessing a rollout, useful questions are concrete: which exact application was evaluated, for whom, with which equipment, and what happens after a positive result? Those questions concern the service around the model as well as the model itself.

Model Current has not independently tested the products or verified the deployment totals in Google’s account. This explainer does not offer individual medical advice. It explains why model availability, screening evidence and diagnosis should remain distinct when describing progress.

Sources

  1. Google Research: TB screening updateOctober 23, 2026 company account supplies the current hook and describes partner deployments and open-model work. Deployment totals, impact estimates and certification claims are not adopted as independently verified facts.
  2. WHO: CAD screening assessment and follow-upJune 11, 2025 primary WHO update: six assessed products, screening age scope and confirmatory testing after a positive screen. Not interpreted as approval of MedGemma or every AI product.
  3. Google: MedGemma model cardCurrent primary documentation, particularly intended use and limitations: development starting point, adaptation, validation, preliminary outputs and no direct clinical diagnosis or treatment use. No benchmark result substituted for local validation.