Truace
Subscribe

Impression-regularized finetuning enables small language models to extract cervical spine fractures from radiology reports

We assessed whether small, open-weight large language models (LLMs) can accurately extract vertebral-level acute cervical spine fracture labels from CT radiology reports while supporting on-premises deployment, and whether impression-regularized finetuning improves accuracy, data efficiency, and temporal stability. In this single-center retrospective study, 8,567 adult cervical spine CT reports (2017-2022) were labeled for acute occipital condyle and C1-C7 fractures with neuroradiologist oversight. Reports from…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2210· cite & verify
Impression-regularized finetuning enables small language models to extract cervical spine fractures from radiology reports
Breast and Cervical Cancer Screening: Barriers and Use Among Specific Populations, Supplement 3 by Casey, S.;Plomer, K.;Tenney, M.. Public domain

In brief

Researchers tested whether small, open-weight language models could pull vertebral-level acute cervical spine fracture labels out of CT radiology reports. They compared standard low-rank adaptation with a version that adds an auxiliary loss based on synthetic, fracture-focused impressions, using 8,567 reports from one center collected between 2017 and 2022.

The approach matters because it suggests that modest models can run inside a hospital's own infrastructure rather than depending on large external systems, which supports privacy and auditability. The evidence comes from a single center and retrospective data, so performance elsewhere and in live clinical workflows is not established by this work.

Main points

  1. Gemma-3-4B macro-F1 rose from 0.885 to 0.909 with the regularized method, and run-to-run variance fell.
  2. The smallest model, Gemma-3-1B, improved from 0.646 to 0.782, showing the largest gains in small models and low-data settings.
  3. On 2018-2022 temporal testing, mean F1 increased from 0.871 to 0.893 with similar or lower variability across time windows.
  4. Remaining errors were mainly linked to ambiguous report language and label-boundary cases.

The gain

A regularized finetuning method raised small open-weight models' accuracy at extracting vertebral-level fracture labels, bringing a 4B model close to a much larger zero-shot model while allowing privacy-preserving local use.

The rundown

The dataset consisted of 8,567 adult cervical spine CT reports labeled for acute occipital condyle and C1-C7 fractures with neuroradiologist oversight. Only the 1,064 reports from 2017 were used for finetuning, while 7,503 reports from 2018-2022 served as a temporal test set.

Models tested included Gemma-3 (1B, 4B, 12B), Llama-3.2-3B, and Phi-3.5-Mini. During training the model samples several synthetic targeted impressions, picks the one most consistent with reference labels, and applies an auxiliary impression loss; at inference, each model does single-pass deterministic JSON extraction.

Baselines were zero-shot prompting, two-pass impression-guided prompting, and standard LoRA. The tuned 4B model reached macro-F1 of 0.909 against 0.913 for zero-shot Gemma-3-27B, with macro-precision 0.929 and macro-recall 0.893.

Sources

  1. Peer-reviewedPLOS Digital Health2026-10-21

The debate