Large language models show potential in MRI protocoling
Large language models can perform on par with radiologists when it comes to selecting the most appropriate MRI protocols for patients, according to a new analysis.
The findings suggest that certain models could help standardize MRI protocol selection, which is typically done manually. This could significantly reduce the time required for manual review and save radiologists valuable time.
“In clinical practice, protocol selection remains largely manual and requires substantial expertise in MRI physics, anatomy and pathology to balance diagnostic accuracy with examination time,” Dr. Zeynep Bendella, with the department of neuroradiology at University Hospital Bonn in Germany, and colleagues write. “This complex process is time-consuming and is often influenced by subjective judgment and institutional standards, which can lead to heterogeneous protocol choices and inconsistent image quality, while incomplete or unnecessarily extended examinations reduce efficiency and may contribute to diagnostic errors or delays.”
Researchers evaluated four open-weight LLMs using 1,200 consecutive MRI referral forms. Each model received the original referral text and selected an institutional imaging protocol and optional sequences from predefined lists. A consensus of two board-certified radiologists served as the reference standard.
Two larger, general purpose models achieved the strongest results. GPT-OSS-120B and Qwen3-235B-A22B-Instruct-2507 each earned 94% protocol accuracy, indicating strong agreement with expert sequence selection. Performance was lower for the medical-focused MedGemma-27B-Text-IT model, which achieved 79% protocol accuracy; Gemma-3-27B-IT had the lowest performance, with 58% accuracy.
The proportion of unacceptable protocol selections also differed considerably. GPT-OSS produced 70 unacceptable selections (5.8%), while Qwen3 produced 71 (5.9%). The rates were higher for MedGemma, at 24.7%, and Gemma3, at 68.3%.
Among unacceptable selections, the most common error involved choosing a protocol that was anatomically appropriate but clinically suboptimal. This accounted for 57.7% to 70.7% of errors across the models. Other mistakes involved selecting the wrong anatomical region or errors limited to optional sequences.
The models also differed in estimated scan duration. GPT-OSS produced durations that were more in line with radiologists' recommendations, while Qwen3 selected shorter protocols. Gemma3 and MedGemma selected protocols with significantly longer scan durations.
While the performances of some of the models are promising, the authors cautioned that the variation in outputs and frequency of unacceptable selections highlight the risks of relying on automated recommendations without clinical oversight.
Read more here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.