MedRise Medical News

Google Releases MedGemma, Open AI Models That Draft Chest X-Ray Reports

Google released new open MedGemma models on July 9, 2025, citing one radiologist who judged 81% of its smallest model's chest X-ray reports good enough to lead to similar patient management, according to Google Research. That review was unblinded.

The models are free to download, adapt and run on a hospital's own servers, so patient data does not have to leave. For students and residents, the useful skill is reading the test behind a number like 81%: who judged, whether they knew the source, and what the model got wrong.

What Google released on July 9, 2025

  • MedGemma 27B Multimodal, a larger model that handles both medical images and text. It joined the earlier MedGemma 4B multimodal and 27B text-only models.
  • MedSigLIP, a lightweight image and text encoder with 400 million parameters. Google said developers at DeepHealth in Massachusetts were exploring it for chest X-ray triage and nodule detection.
  • All sit in Google's Health AI Developer Foundations (HAI-DEF), which Google describes as starting points for developers, not finished clinical tools.

The 81% figure, read like a reviewer

What Google reported in July 2025, and the question each number should trigger
Google's claimWhat it meansAsk
81% of MedGemma 4B chest X-ray reports would lead to similar managementAbout 1 report in 5 would notWhat went wrong in that fifth?
The review was unblindedThe reviewer knew which reports came from the modelWould a blinded reader agree?
One US board-certified radiologist judged themNo second reader, so no measure of agreementHow many readers, and did they agree with each other?
MedGemma 4B scored 64.4% on MedQAA multiple-choice exam benchmark, not a written reportHow does it do on free-text reports from real patients?

What has happened since (updated November 17, 2026)

  • December 19, 2025: Google introduced MedASR, an open speech-to-text model for medical dictation, on its HAI-DEF blog.
  • January 13, 2026: MedGemma 1.5 4B added 3D CT and MRI volumes, whole-slide pathology, chest X-ray time series, anatomy localization and lab-report extraction. Against MedGemma 1, Google reported MRI finding classification of 65% vs 51%, lab-report extraction of 78% vs 60% (macro F1) and MedQA of 69% vs 64%. Google said MedASR made 58% fewer errors than Whisper large-v3 on chest X-ray dictation (5.2% vs 12.5% word error rate). Google said the models remain free for research and commercial use.
  • March 26, 2026: Google announced the winners of its Kaggle MedGemma Impact Challenge, which drew over 850 teams. The winners were prototypes, such as an on-device tuberculosis screening workflow that combines chest X-rays and cough audio.
  • October 2026: Google researchers published the MedGemma and MedSigLIP work in Nature Medicine. On out-of-distribution tests, they reported improvements of 15.5% to 18.1% over the base Gemma 3 models in chest X-ray finding classification.

Blinded, independent tests found the limits

Studies indexed in PubMed in 2026 tested MedGemma outside Google, and they are harder graders than an unblinded single reader.

  • Hallucinated findings. In a European Radiology study of 478 emergency patients, three thoracic radiologists blindly rated chest X-ray reports from five medical AI models, including MedGemma, against radiologists' own reports. The best model was not open source. The other four, MedGemma among them, hallucinated findings in 5.4% to 17.4% of reports, against 0.1% for radiologists, and were rated clinically acceptable in 41.1% to 71.4% of cases, against 74.3%.
  • Missing context, missed disease. A Pediatric Radiology review from Johannesburg reported critical failures in detecting pericardial effusion and tuberculosis when MedGemma was not given explicit clinical context.
  • Local tuning matters. A Turkish spine MRI study fine-tuned MedGemma 4B on local reports (about 10,000 in the study). Agreement with reference impressions rose from 17.7% to 67.1%, although those references came from another AI model, not a radiologist.
  • Exams flatter models. On the multiple-choice ReXVQA benchmark, MedGemma scored 83.84% on a 200-case sample, above the best of three radiology residents (77.27%). Choosing an answer is not the same as writing a safe report.

The fine print in Google's license

  • Google's HAI-DEF FAQ says commercial use is allowed under its terms, and that running the open weights on your own servers keeps data inside your environment.
  • The terms of use require developers to seek health regulatory authorization where it applies, and anyone who redistributes the models or a modified version must pass the use restrictions on.
  • Google's MedGemma page says the model "isn't yet clinical-grade." Google's 2026 post says outputs are preliminary and need independent verification.

What to do with this

  • Run the four-question test on any AI accuracy claim: blinded or not, how many readers, which patients, compared with what. Our guide to AI in medicine lists more questions.
  • Give the clinical context. When you test a model on an image, include the history. The pediatric evaluation shows what goes missing without it.
  • Read a MedQA score as a quiz score. 64% to 69% says little about how the model handles a real film.
  • Turn it into a project. MedGemma 1.5 4B is small enough to run offline. A blinded, two-reader evaluation on local cases is a realistic student research project. The MedRise research course shows how to scope one.

The models got better at reading scans in 2026. The independent tests got better at catching their mistakes, and they are the ones to read first.

Sources

MedRise Medical News summarizes public announcements and research for education only. It is not medical, legal or immigration advice.

Book a free strategy call

Ten minutes to look at your exam date, your scores and what is getting in the way.