Radiology Report Generation

Image encoder plus LLM generating radiology findings.

multimodal LLM

3named models

180candidates

2deployments

1in Indonesia

Named Models

MedGemma 4BGoogleReleasednot in the notesRuns onnot in the notesClassmultimodal LLM

4B multimodal medical model built on Gemma 3 4B, used for radiology report generation and medical QA. Dermatology images among its inputs is unverified in the notes.Commercial use: not stated; Jakarta hosting: unknown; Indonesian support: not stated

MAIRA-2Microsoft AIReleasedJun 2024Runs onnot in the notesClassmultimodal LLM7B model that generates grounded chest X-ray findings; Microsoft reported state of the art on MIMIC-CXR findings generation in June 2024. Research use only.Commercial use: non-commercial; Jakarta hosting: unknown; Indonesian support: not stated

CheXagentStanford UniversityReleasednot in the notesRuns onnot in the notesClassmultimodal LLM

Radiology report generation model attributed to Stanford; unverified in the notes.Commercial use: not stated; Jakarta hosting: unknown; Indonesian support: not stated

Where this comes from

Models the research names for this use case. Research, with sources on each model record.

Who Runs It

production, 1planned, 1

GleamerCountry: FranceRadiology reporting with Med-PaLM and Gemini on Vertex AIVendor: Google CloudproductionDate–More than 35 million exams a year (claim)Evidence: company statementConfidence: lowSource: cloud.google.com

Kementerian Kesehatan (Kemenkes), PusdatinCountry: IndonesiaCentral AI reading of images (e.g. MRI) from primary care facilities linked to SatuSehat (model not stated)Vendor: not statedplannedDate19 Aug 2026Images from primary facilities would get a first AI reading centrally and be sent back to facilities short of doctors; nothing measuredEvidence: government statementConfidence: mediumSource: rctiplus.com

Where this comes from

Documented uses by named organisations. The evidence column says who made the claim.

Candidate Models

Artificial Analysis Intelligence Index: AA leaderboard, 205 current models

A general benchmark for this model class, not a test on this use case. Composite intelligence score across AA evaluations. index. Higher is better. Showing 10 of 17.

Claude Opus 5.5Anthropic independent58

Claude Sonnet 5.5Anthropic independent56

Claude Fable 5.1Anthropic independent53

GPT-6 AstraOpenAI independent53

GPT-6.1 SolOpenAI independent52

Muse Spark 1.3Meta independent48

Grok 4.7SpaceXAI (xAI) independent46

MiMo-V2.6-ProXiaomi independent46

Qwen3.8 MaxAlibaba (Qwen) independent45

Kimi K3Moonshot AI independent44

050100

Apertus v1.5 70BSwiss AI (EPFL, ETH Zurich, CSCS), multimodal LLMReleasedSep 2026open weightsIndonesian statedhas scoreslast 18 months

Muse Glimmer 30BMeta, multimodal LLMReleasedAug 2026open weightsIndonesian statedhas scoreslast 18 months

Mistral Medium 3.5 (v26.04)Mistral AI, multimodal LLMReleasedApr 2026open weightsIndonesian statedhas scoreslast 18 months

Gemma 4 31BGoogle, Google DeepMind, multimodal LLMReleasedApr 2026open weightsIndonesian statedhas scoreslast 18 months

Qwen3.6-27BAlibaba (Qwen), multimodal LLMReleasedApr 2026open weightsIndonesian statedhas scoreslast 18 months

Qwen3.5-4BAlibaba (Qwen), multimodal LLMReleasedMar 2026open weightsruns on edgehas scoreslast 18 months

Qwen3.5-122B-A10BAlibaba (Qwen), multimodal LLMReleasedFeb 2026open weightsIndonesian statedhas scoreslast 18 months

MERaLiON-2-3BA*STAR (Institute for Infocomm Research, I2R), multimodal LLMReleasedMay 2025open weightsIndonesian statedhas scoreslast 18 months

MERaLiON-2-10BA*STAR (Institute for Infocomm Research, I2R), multimodal LLMReleasedMay 2025open weightsIndonesian statedhas scoreslast 18 months

MiMo-V2.6-FlashXiaomi, multimodal LLMReleasedSep 2026open weightshas scoreslast 18 months

Showing 10 of 180. All models of this class

Where this comes from

Derived, not researched: models in the directory whose class matches a class this use case uses, excluding family records and models shut down or deprecated. Ranked by how many fit facts they meet, then by release date. Each fit fact is read from the model record.

Use Case

Industry: healthcare and pharma
Model classes: multimodal LLM
Benchmarks or datasets: RadGraph F1 MIMIC-CXR CheXpert
Regulatory notes: Surat Edaran Menkominfo 9/2023 Etika Kecerdasan Artifisial, Kementerian Komunikasi dan Informatika (now Kementerian Komunikasi dan Digital)Radiology report generation is an AI-based service run by an electronic system operator, and the circular asks such operators to build and use AI on the values of inclusivity, transparency, humanity and security.
UU 27/2022 Pelindungan Data Pribadi, DPR and President of the Republic of Indonesia (administered by Kementerian Komunikasi dan Digital)Radiology report generation processes health data of patients, which the law treats as a specific category of personal data, so the controller needs a lawful basis, must protect the data and must honour data subject rights.
UU 17/2023 Kesehatan, DPR and President of the Republic of Indonesia (administered by Kementerian Kesehatan)Radiology report generation is used in delivering health services or handling patient records, which this law governs through rules on medical records, health information systems, health technology and medical devices.
EU AI Act, Regulation (EU) 2024/1689 (international), European UnionRadiology report generation is software that may qualify as a medical device or a safety component of one; the EU AI Act treats such AI as high-risk when the product is placed on the EU market.

Where this comes from

From the industry tracks T08 to T10 of the research notes.

R&D Topics

Links to R&D topics are awaiting review.

Where this comes from

Which R&D topics would most change this use case. Drafted by Claude from the topic definitions and the use case; proposed until approved.

Incidents

Google Healthcare AI Model Med‑Gemini Allegedly Produces Non‑Existent 'Basilar Ganglia' Term in Published Output2024-05-06, safetyGoogle’s Med‑Gemini healthcare AI reportedly produced the non‑existent term "basilar ganglia" in public launch materials, conflating two distinct brain structures.

Where this comes from

Documented failures, harms or misuse in this kind of use, Indonesia first, then the region, then well-known cases elsewhere (AI Incident Database, OECD AI incidents monitor, reputable news). Not a claim about any deployment listed here.

Opportunities

Opportunities are awaiting review.

Where this comes from

Gaps found by rule in what the directory holds for this use case. Each is a proposal for review.

Links

IndustryHealthcare40 use cases, 66 deployments

Use CaseRadiology Report GenerationHealthcare

Models

MAIRA-2Microsoft AI, multimodal LLM

MedGemma 4BGoogle, multimodal LLM

CheXagentStanford University, multimodal LLM

Who Runs It

Kementerian Kesehatan (Kemenkes), PusdatinPlanned, 19 Aug 2026 GleamerIn production

Graph

Models 3

MedGemma 4BGoogle, multimodal LLM

MAIRA-2Microsoft AI, multimodal LLM

CheXagentStanford University, multimodal LLM

Deployments 2

GleamerRadiology reporting with Med-PaLM and Gemini on Vertex AI, France Kementerian Kesehatan (Kemenkes), PusdatinCentral AI reading of images (e.g. MRI) from primary care facilities linked to SatuSehat (model not stated), Indonesia

Sources

huggingface.co/microsoft/maira-2 https://huggingface.co/microsoft/maira-2
MedGemma Technical Report arxiv.org https://arxiv.org/html/2507.05201v2
MAIRA-2: Grounded Radiology Report Generation - Microsoft Research microsoft.com https://www.microsoft.com/en-us/research/publication/maira-2-grounded-radiology-report-generation/

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Radiology Report Generation", https://aurion.semai.id/usecases/hc-radiology-report-generation/, records of October 2026.