Structuring Turkish lumbar and cervical spine MRI reports with a fine-tuned compact language model

Source article: TurkRadLLM: A Compact MedGemma-Based Model for Structuring Turkish Lumbar and Cervical Spine MRI Reports-A Multi-reader Validation Study

To develop and validate a compact, locally deployable MedGemma-based model converting unstructured Turkish spine MRI reports into structured six-field output. In this retrospective single-institution study, 5210 lumbar and 5017 cervical Turkish spine MRI reports were used to fine-tune MedGemma 4B IT with QLoRA. Silver-standard labels from a proprietary large language model served as training targets and as the reference for automated fidelity analyses, but not as the clinical reference standard. Evaluation compr…

By the Truvace deskOctober 21, 20262 min read1 sourcePeer-reviewedTRV-2026-2087· cite & verify

Trace impact readingContested

P 67

Both sides are scored from claims and sources, not community votes.

G 65

In brief

Researchers fine-tuned a small MedGemma 4B model on thousands of Turkish lumbar and cervical spine MRI reports to convert free text into a six-field structured format. Readers reviewed 180 real reports mixed with 20 adversarial cases, and every output was valid JSON that readers accepted as a group.

The result suggests a modest model can run locally on consumer hardware for report structuring, which matters where data privacy or compute is limited. However, training labels came from another language model rather than a clinical reference standard, and one reader found negative-finding lists incomplete often, so clinical reliability is not settled.

Main points

The model was fine-tuned with QLoRA on 5210 lumbar and 5017 cervical Turkish MRI reports from a single institution.
Agreement with the silver-label impression rose from 17.7% to 67.1% on 1025 held-out reports after fine-tuning.
Urgent-flag sensitivity against the silver reference rose from 18.8% to 62.5%, versus 50.0% for a much larger non-fine-tuned model.
A rule-based baseline reached 99.0% on body region, showing simple methods suffice for some fields.
Deployment needed 14.75 GB of GPU memory on a consumer card.

The gain

A fine-tuned compact MedGemma model produced valid structured output from Turkish spine MRI reports that readers accepted, and it ran on a single consumer GPU card.

The problem

Negative-finding completeness was reader-dependent, with one reader judging negative-finding lists incomplete in 57.8% of reports, and fidelity to the silver-label impression reached only 67.1%.

The rundown

The study was retrospective and single-institution. Training targets were silver-standard labels from a proprietary large language model, which were also the reference for automated fidelity analysis but not the clinical reference standard.

Validation combined multi-reader scoring of 180 authentic reports interleaved with 20 adversarial cases, a second round where readers scored the silver labels themselves, and comparison on the full held-out set against a rule-based baseline, the base model, and a much larger model.

When readers scored the silver labels themselves, the fine-tuned model was rated higher on negative findings and lower on the impression. The authors conclude the gain lay in content-level fidelity rather than output validity.

What this doesn’t fix

Labels came from a proprietary language model rather than a clinical reference standard, the study was single-institution, and negative-finding completeness varied by reader.