LLMs can pick the right MRI protocol.
The report is still the bottleneck.

The study: four LLMs, 1,201 real referrals, one task

A team led by researchers at University Hospital Bonn, Germany, published the results in European Radiology, as first reported by Radiology Business. The team tested four open-weight (locally deployable, privacy-compliant) LLMs on 1,201 consecutive MRI referral forms, feeding each model the unaltered referral text and asking it to choose an institutional protocol and any optional sequences from the department's predefined list. A consensus of two board-certified radiologists served as the reference standard.

The two largest, general-purpose models — GPT-OSS-120B and Qwen3-235B-A22B-Instruct-2507 — each matched that expert consensus on 94% of cases, with a Jaccard overlap of 0.96 against the reference protocol and sequence set. Notably, the smaller, medically fine-tuned MedGemma-27B-Text-IT model trailed at 79% accuracy, and the smallest general model, Gemma-3-27B-IT, managed only 58%. Bigger, broadly-trained models beat a narrower medical model at this task — a pattern the authors tie to stronger general reasoning and more consistent instruction-following, not domain fine-tuning.

Where it held up, and where it didn't

Accuracy alone understates the gap between models. The study also counted outright "unacceptable" selections — protocols a radiologist would reject, not just quibble with. GPT-OSS-120B produced 70 unacceptable picks out of 1,201 cases (5.8%); Qwen3 produced 71 (5.9%). MedGemma's unacceptable rate jumped to 24.7%, and Gemma-3's reached 68.3% — meaning the weakest model got more than two in three cases wrong by the study's own bar.

Even among the strong models, errors had a pattern: the single most common mistake, across most models, was choosing a protocol that was anatomically correct but clinically suboptimal — the right body part, but not the acquisition tailored to the actual clinical question — accounting for roughly 58–71% of all errors depending on the model. Wrong-region and sequence-level mistakes made up the rest. On timing, the top model's chosen protocols ran close to the radiologists' own choice (GPT-OSS matched expert scan duration; Qwen3's picks ran slightly shorter, about 19 minutes 15 seconds versus 19 minutes 58 seconds on average) — evidence the better models aren't just accurate on paper, they're not quietly padding scan time either.

Figures: Bendella et al., European Radiology (2026), 1,201 referrals, reference = consensus of two board-certified radiologists.

The authors' own caveat

The researchers were careful not to oversell the result. They noted the dataset came from a single institution with its own protocol library and ground truth, that 14.2% of initially retrieved referrals had to be excluded for incomplete clinical information, that each model only ran once per case (so run-to-run variability wasn't measured), and that the reference radiologists partially overlapped with those used for concordance scoring. Their conclusion: the variation across models, and the frequency of unacceptable selections even among the better ones, show the real risk of leaning on automated protocol recommendations without a human checking the output.

That caveat matters for how the result should be read. This is not "LLMs can run MRI protocoling unsupervised." It's "the best current open-weight LLMs are close enough to expert-level on a closed-list classification task that it's now a credible decision-support tool" — provided a radiologist stays in the loop for the harder cases.

Why protocoling is the easy end of the workflow

It's worth being precise about what this study actually automates. MRI protocol selection is a closed-list classification problem: the model reads a referral and picks from a fixed, finite menu of institutional protocols and sequences. There's a right answer, a known answer set, and a straightforward way to score agreement. That's exactly the kind of task LLMs are good at — and it's the same reason two other, distinct LLM-automation studies in radiology have shown strong results recently: LLMs writing more complete clinical indications for imaging orders than the ordering clinicians themselves, and LLMs catching quality-control errors in finished reports faster and more consistently than manual review. Protocoling, order-writing, and QC auditing sit at the edges of the radiology workflow — structured inputs or structured checks, not open-ended report generation.

Structured CT reporting — the actual diagnostic narrative a radiologist signs — is a different kind of problem. It's open-ended: the model has to reason over image findings, synthesize them into prose with the right level of certainty, flag incidental findings, and produce something a referring clinician will act on. It carries direct diagnostic liability in a way protocol selection never will, because an unacceptable protocol choice delays or recaptures a scan, while an unacceptable report finding can delay a diagnosis. That gap in stakes and structure is a large part of why even purely human-authored structured reporting has been slow to spread: a 2024 European Radiology survey of 200 oncologic-imaging radiologists across 51 countries found only 57% used structured reporting in daily practice at all, and just 51% among European respondents — years into the push for it, without AI in the mix.

Put together, the picture across all three LLM-automation studies is consistent: the steps around the report — what to order, what protocol to run, whether the finished report is clean — are converging on automation faster than the report itself. The report remains the place where a radiologist's judgment is hardest to replace and where oversight matters most.

Where xAID fits

This is the problem xAID's foundation-model approach to CT reporting is built around: rather than treating report drafting as a classification task, the model generates a full structured report directly from the imaging data, xAID's in-house radiologist reviews every preliminary, and the report reaches the client ready-to-sign. Protocol selection and order QC can be checklist problems for an LLM; a diagnostic CT report can't be — which is why the review step doesn't disappear as the surrounding workflow automates, it becomes the whole point.

Frequently asked questions

Source: Bendella et al., "Accuracy and reliability of privacy-compliant large language models for MRI protocol selection in radiological workflows," European Radiology (2026), DOI: 10.1007/s00330-026-13039-9, as reported by Radiology Business. Structured-reporting adoption figures from a 2024 European Radiology survey of ESOI members. Figures are rounded as reported.