Key takeaway: MedGemma is a family of medical AI models that developers can adapt for imaging and text applications. Its research results are promising, but the model is not ready to make independent clinical decisions. Every proposed application needs appropriate validation, privacy protections, oversight and regulatory review. [1, 2]
Could an AI model help radiographers learn anatomy, explore a teaching image, or find a departmental protocol? Google’s MedGemma brings medical images and language together, opening up possibilities for education and research. But a fluent answer does not guarantee a correct one, and interpreting images, assessing image quality and communicating with patients are different tasks.
What is MedGemma?
MedGemma is a collection of open-weight models built from Google’s Gemma family and adapted for medical text and image comprehension. Open-weight means developers can obtain model parameters under the applicable terms and adapt them; it does not mean a finished application is automatically safe, clinically approved or free to operate. [1, 2]
MedGemma 1.5 4B accepts images and text and generates text. It has expanded support for CT and MRI volumes, images taken at different times, anatomical localization and medical-document understanding. Earlier MedGemma 1 models include 4B and 27B variants; the 27B family includes text-only and multimodal versions. The text-only model cannot interpret an image. Model size alone does not determine which is best for a particular task. [1, 2]
Google also states the intended use narrowly. MedGemma is “intended to be used as a starting point that enables more efficient development of downstream healthcare applications,” it is aimed at developers rather than at end users, and the model card says its outputs “are not intended to directly inform clinical diagnosis, patient management decisions, treatment recommendations, or any other direct clinical practice applications.” Treat that as a design boundary rather than a legal formality: it caps what any benchmark score below can mean. [1]
Where might it help radiographers?
- Education: Help explain anatomy, terminology and positioning concepts using verified references and approved teaching materials.
- Image-quality teaching: Generate discussion points about rotation, inspiration and anatomical coverage, then check them against a structured image critique routine. A convincing explanation is not an image-acceptance decision.
- Finding information: Help locate or summarize approved protocols or educational material, with staff checking that the source is current.
- Research: Explore questions involving medical images, longitudinal studies or documents under appropriate research and privacy oversight.
- Document assistance: Developers can investigate extracting information from laboratory reports or records. Outputs still require verification and an authorized workflow.
These are possible uses, not evidence that MedGemma has been validated for each one. A model trained to answer questions about an image is not automatically reliable at assessing positioning, recommending a repeat exposure, or detecting a particular pathology.
A practical teaching example
A student asks, “What should I check when evaluating a chest radiograph?” using an approved, appropriately de-identified teaching image. The model might mention rotation, inspiration, exposure and anatomical coverage. The student can then compare the answer with course material and an educator’s assessment. The AI may miss a problem or describe one that is not present, so the student should not use it to decide whether a patient image is acceptable.
What does the research show?
Google’s model card reports results for specific datasets and tasks. On an internally curated CT dataset, MedGemma 1.5 4B achieved 61.1% macro accuracy across seven conditions, compared with 58.2% for MedGemma 1 4B. On an internal MRI dataset, the reported figures were 64.7% and 51.3% across ten conditions. On CT-RATE validation, MedGemma 1.5 achieved 27.0% macro F1 across 18 thoracic findings. [1]
These figures measure different things: accuracy and F1 are not interchangeable, and none represents the accuracy of a complete clinical diagnosis. The internal CT and MRI evaluation samples were balanced by condition, so results should not be assumed to reflect routine clinical prevalence. Performance may differ with new scanners, protocols, populations or institutions. [1]
MedGemma 1.5 also improved on some tasks but declined on others. For the SLAKE visual question-answering benchmark, its tokenized F1 was 59.7, compared with 72.3 for MedGemma 1 4B. A newer model is therefore not necessarily better for every job. Google notes that version 1.5 was less optimized for this question-and-answer format; the benchmark alone does not establish why the change occurred. [1]
Recognizing a finding and locating it are different
In an independent chest X-ray study, Gosai and colleagues used a grid-based task to ask models where a specified abnormality was most prominent. Across nine findings, the tested MedGemma model had a 17.7% localization hit rate, compared with 59.9% for a task-specific convolutional neural network and 80.1% for a radiologist benchmark. The result applies to the tested model, dataset and method; it should not be generalized to all MedGemma versions or clinical tasks. [3]
Google separately reported 38.0% intersection over union for MedGemma 1.5 4B on an anatomical bounding-box task. That measures a different task and outcome from the pathology-localization study, so the numbers cannot be compared directly. [1, 3]
Is this research, available software, or approved clinical practice?
Those are three different questions, and a single conversation can slide between them without anyone noticing. The table separates them. Every row has a narrow meaning, and the practical question for each is the same: what decision is this output allowed to influence?
| What you might hear | Status of the evidence | What it establishes | What it does not establish |
|---|---|---|---|
| “MedGemma scores 61.1% on CT” | Human research — retrospective evaluation on an internally curated dataset | The tested version's macro accuracy across seven conditions, under that dataset's sampling | Accuracy on your scanner, protocol and patient mix; examination-level diagnostic accuracy |
| “MedGemma localizes pathology at 17.7%” | Human research — one independent study, grid-based localization task | That model's hit rate on CheXlocalize beside named comparators | That every MedGemma version behaves alike, or that localization equals clinical usefulness |
| “It is open-weight, so we can run it locally” | Clinically available software component, evolving | Weights can be obtained and adapted within the terms of use | Security, privacy compliance, device licensing or fitness for a specific task |
| “The DAP AI standards apply from January 1, 2027” | Regulatory — accreditation standards for BC diagnostic facilities | Facility accreditation requirements, including committee monitoring of interpretative AI | Device authorization, or that any specific AI product is approved for clinical use |
| “Generative AI has to comply with privacy law” | Regulatory — BC privacy statutes and regulator guidance | How personal information must be handled, and that AI sits inside existing law | Whether a tool is accurate, or that a particular use is approved by your employer |
Read together, the cited material supports a research-stage and retrospective claim, not a clinical-practice claim. No source in this article reports a prospective deployment, an outcome improvement in a live department, or a regulator authorizing MedGemma for a specific clinical task. [1, 2, 3, 7]
Potential advantages
- Images and language together: An image and question can produce a conversational response useful for research or learning.
- Adaptability: Researchers can investigate focused applications using the available weights and terms of use.
- Flexible deployment: Local deployment may be possible with suitable hardware and software. This can support institutional control but does not automatically satisfy privacy, security or clinical requirements.
- Broader research capabilities: The newer model supports selected volumetric, longitudinal and document tasks; the suitability of each application still requires testing. [1, 2]
Limitations and risks
- Confident errors: The model may miss a finding, invent one or describe its location incorrectly. Human-sounding text is not evidence of correctness.
- Task and population differences: Benchmark results do not prove performance on a local patient population or a different workflow.
- Image-quality decisions: A model would need specific validation for the projections and local criteria involved. Repeat-exposure decisions require professional judgment and radiation-protection considerations.
- Automation bias: Users may defer to an answer that sounds certain. Staff need ways to question results, report errors and retain independent judgment.
- Privacy: Patient identifiers may appear in image pixels, DICOM metadata and associated records. Do not enter patient information into an unapproved AI tool; follow local policies and approved data-handling procedures.
- Implementation costs: A useful application can require computing resources, integration, testing, security, maintenance and monitoring. Local deployment alone does not establish compliance.
- Unequal performance: Evaluate whether error rates differ across relevant patient groups, equipment and settings. [1]
Learning exercise: reading an AI claim like a radiographer
Six items. Decide first, then read why the alternatives fail — the reasoning is the transferable part, not the answer. Every figure is quoted from the reference list, and the arithmetic in item 3 is shown so you can redo it.
1. The balanced-dataset trap
Claim: “MedGemma 1.5 4B reached 61.1% macro accuracy on CT, so it gets roughly 61 out of every 100 CT examinations right.”
- Correct — 61.1% is an accuracy figure, so it describes examination-level accuracy.
- Partly right, but the figure is tied to seven conditions on a condition-balanced internal dataset, so it cannot be read as examination-level diagnostic accuracy.
- Wrong, because the reported CT figure is 27.0, not 61.1.
- Wrong, because Google does not publish results for internal datasets.
Best response: B.
- A assumes an accuracy metric maps onto examinations. The model card labels CT Dataset 1 as internally curated and balanced by condition, so the prevalence in that sample is not the prevalence in a department.
- C mixes up two different rows: 61.1 is macro accuracy on CT Dataset 1, while 27.0 is macro F1 on the CT-RATE validation set. Different dataset, different metric. [1]
- D is factually wrong. Internal-dataset results are published; the limitation is generalizability, not secrecy. [1]
2. Localization is not detection
Claim: “MedGemma localizes chest radiograph abnormalities better than a task-specific CNN — Google measured 38.0% IoU.”
- Correct — 38.0% is higher than the CNN's 59.9%.
- The 38.0% is anatomical bounding-box detection on Chest ImaGenome; the independent pathology-localization study reported 17.7% for MedGemma against 59.9% for a CNN, so no superiority claim follows.
- Both figures are comparable, but the study is now out of date.
- Localization and detection are the same task, so the comparison is valid.
Best response: B.
- A fails on arithmetic as well as on task: 38.0 is less than 59.9, and the two numbers come from different tasks.
- C accepts comparability that does not exist — anatomy boxes and pathology hit rates are different outcomes. [1, 3]
- D is the error this exercise exists to catch. Boxing an anatomical structure is not the same as pointing to where a named abnormality predominates.
3. Worked arithmetic: how big is the gap?
The independent study reported a 17.7% localization hit rate for the tested MedGemma model and an 80.1% radiologist benchmark.
- Absolute difference: 80.1 − 17.7 = 62.4 percentage points (point estimates, same dataset and task).
- Relative level: 17.7 ÷ 80.1 = 0.221, so MedGemma reached about 22.1% of the radiologist benchmark.
- Unit check: a difference between two percentages is in percentage points, not percent; the ratio is dimensionless and is reported as a percentage of the benchmark.
Which statement is correct?
- MedGemma performed 62.4% as well as radiologists.
- MedGemma performed at about 22.1% of the radiologist benchmark, a gap of 62.4 percentage points.
- The correct comparison is 80.1 ÷ 17.7 = 4.5, so MedGemma performed at 4.5% of the benchmark.
- The gap is 6.24 percentage points.
Best response: B.
- A reports a percentage-point difference as a performance share — the commonest misuse of these figures.
- C divides in the wrong direction and then mislabels the result; 4.5 is a dimensionless multiple, roughly how many times higher the benchmark is.
- D is an arithmetic slip.
Note on scope: these are point estimates from one study and one task. The comparison supports “far below the radiologist benchmark on this task”, not a general statement about MedGemma's clinical value. [3]
4. A newer version is not automatically better
Claim: “MedGemma 1.5 is newer than MedGemma 1 4B, so it is at least as good at every task.”
- Correct — later versions dominate earlier ones.
- Incorrect — on SLAKE radiology tokenized F1, MedGemma 1.5 4B scored 59.7 against 72.3 for MedGemma 1 4B, and the model card notes that 1.5 was less optimized for that question-and-answer format.
- Incorrect — but only because 1.5 scored higher on every published row.
- Correct, because a higher version number implies a strictly larger training set.
Best response: B.
- A and D are assumptions about versioning, not findings. The model card reports gains in 3D and longitudinal tasks alongside this regression.
- C contradicts the published table. [1]
- The lesson for a department: choose a version against your own task, not against its release date.
5. What an accreditation standard actually does
Claim: “BC's DAP artificial intelligence standards take effect on January 1, 2027, so interpretative AI is approved for clinical use in BC from that date.”
- Correct — accreditation standards authorize devices.
- Incorrect — DAP version 2.1 is an accreditation standard for diagnostic facilities that expects an AI optimization committee and an annual review of interpretative AI; device authorization is a separate question.
- Incorrect — the standards took effect in 2026.
- Correct, and it applies to every AI tool in every facility.
Best response: B.
- A confuses the accreditor with the device regulator. Standards describe what a facility must do; they do not license a product.
- C confuses two dates in the same document: published May 20, 2026, effective January 1, 2027. [7]
- D is the kind of blanket claim this article avoids. The standards themselves tie requirements to whether interpretative AI is used, and applicability depends on facility and use.
6. Which privacy law is in play?
Scenario: Two technologists each consider pasting a report into a general-purpose chatbot. One works at a BC hospital; the other works in a privately owned clinic.
- PIPA applies to both, because both are in British Columbia.
- FIPPA applies to the health authority; PIPA applies to most private clinics and independent practices.
- PIPEDA applies to both.
- Neither applies if the information has been de-identified.
Best response: B.
- A treats one statute as province-wide. Public bodies such as health authorities and most hospitals rely on FIPPA, a separate authorization that the Commissioner's AI guidance explicitly does not cover in its PIPA analysis. [6]
- C names the federal private-sector statute rather than the BC framework described by the cited guidance.
- D is the most dangerous answer. Identifiers can persist in image pixels, DICOM metadata and associated records, and employer policy governs handling even where de-identification is claimed. [4, 5, 6]
- In both cases the first check is the same: is this specific tool approved for this specific use by your employer?
What about British Columbia?
Rules depend on the intended use, the organization and the professional role, so no single sentence covers every BC workplace. Three separate questions get confused with each other here: privacy law, accreditation, and professional regulation.
Privacy. Fraser Health states that it may use artificial intelligence models both in care and in planning or improving services, that it does so “in accordance with applicable Canadian privacy laws and after performing detailed assessments to make sure privacy, security and other risks are appropriately addressed,” and that identifying information is removed under a strict protocol before its models are trained. [4] That is a statement about the organization's own process. It is not a standing permission for an employee to paste an image or a report into a general-purpose chatbot, and it does not describe every BC employer.
BC's Information and Privacy Commissioner is explicit that generative AI does not sit outside existing law: organizations that develop, provide or use generative AI “are obligated to ensure that their activities comply with applicable privacy laws and regulations in Canada.” [5] Which statute applies depends on who holds the information. Health authorities and most hospitals are public bodies governed by the Freedom of Information and Protection of Privacy Act (FIPPA), while most privately owned clinics and independent practices fall under the Personal Information Protection Act (PIPA); the Commissioner's health-sector AI guidance treats those as separate analyses rather than interchangeable ones. [6] Before entering any patient information into an AI tool, confirm that the specific tool and that specific use are approved by your employer, and use approved systems only.
Accreditation. BC's Diagnostic Accreditation Program has published version 2.1 of its artificial intelligence standards, published May 20, 2026 and effective January 1, 2027. For interpretative AI it requires that a committee be formed to review and monitor the use of interpretative AI (AI4.1.1), with an annual review of accuracy auditing, quality assurance, potential bias and effects on clinical decision-making (AI4.1.2). The suggested membership — users such as physicians, quality assurance personnel such as physicists, people with AI expertise, legal and risk management representatives — appears in the standard's guidance as a recommendation, not as a requirement. The same standards state that interpretative AI must support or supplement interpretation rather than replace comprehensive evaluation by a physician, and that clinical management of an AI-flagged abnormality is determined by a physician. Applicability depends on the facility and the specific use, so consult the DAP and local leadership rather than assuming every tool carries identical requirements. [7]
Professional regulation. Those are accreditation standards for facilities. They are not a device authorization, and they do not approve any AI product for clinical use. Radiographers should work within their scope of practice, local policy and applicable requirements. CAMRT-BC distinguishes medical imaging technologists from radiation therapists and states that radiation therapists are scheduled to enter CPSBC regulation on November 29, 2027; this article therefore makes no blanket claim that the same regulatory pathway applies to every medical radiation technology discipline in BC. [8]
Questions to ask before implementation
- What specific problem is the application intended to solve, and what decisions will users make from its output?
- Has the finished system been validated for that purpose, local equipment, patient groups and workflow?
- Who reviews its output, remains accountable and responds to errors?
- Are privacy, security, data handling and use of patient information approved?
- What device, accreditation, institutional and jurisdictional requirements apply?
- How will performance and potential bias be monitored as software, equipment and populations change?
The takeaway
MedGemma could support radiography education, research and carefully designed workflow tools. Its ability to combine images with language is useful to explore, but research benchmarks and fluent answers do not establish clinical readiness. The right question is not simply whether AI can answer: it is whether a specific application has been validated, approved and supervised for the task in front of it.
Keep learning
MedGemma is one narrow example of a wider question: how much of an image-acceptance or interpretive judgement can be delegated, and to what. These pages build the comparison:
- Image critique and evaluation — the structured judgement an AI explanation is not performing.
- Flat-panel detectors, DQE and the deviation index — how acquisition and processing variables show up as measurable image-quality numbers.
- Photon-counting CT explained — a worked example of separating a hardware change from a software one, and CT from projection radiography.
Next step: test the underlying physics and image-quality material in the free 25-question diagnostics, or work the numbers yourself with the radiography calculators. For how this site handles authorship and review, see the editorial policy.
References
- Google. MedGemma 1.5 model card. Model variants, intended use, limitations, evaluation methods and benchmark results.
- Sellergren A, Gao C, Mahvar F, et al. MedGemma 1.5 Technical Report. arXiv:2604.05081. 2026. Preprint.
- Gosai A, Kavishwar A, McNamara SL, et al. Beyond Diagnosis: Evaluating Multimodal LLMs for Pathology Localization in Chest Radiographs. Proceedings of the 5th Machine Learning for Health (ML4H) Symposium. 2025. arXiv:2509.18015.
- Fraser Health. Artificial intelligence and privacy. Public statement on AI use, privacy assessments and de-identification.
- Office of the Information and Privacy Commissioner for British Columbia. Principles for responsible, trustworthy and privacy-protective generative AI technologies. Guidance document.
- Office of the Information and Privacy Commissioner for British Columbia. PIPA and AI scribes: best practices for healthcare organizations in BC. Guidance document; notes that BC public bodies rely on FIPPA rather than PIPA.
- Diagnostic Accreditation Program, College of Physicians and Surgeons of British Columbia. Accreditation Standards: Artificial Intelligence, version 2.1. Document ID 23849; published 2026-05-20, effective 2027-01-01.
- CAMRT-BC. Regulatory college updates. Timeline for regulation of radiation therapists by CPSBC.
Educational article researched November 10, 2026. AI-assisted writing; no qualified clinical or legal review is claimed. Follow current local policy and professional guidance.