Small AI models learn to spot cervical spine fractures in radiology reports

Small, open-weight language models can accurately extract acute cervical spine fracture labels from CT radiology reports, and a new finetuning method called impression-regularized low-rank adaptation makes them markedly better at it, according to a single-center retrospective study published in PLOS Digital Health.

The research team, led by Zixuan Hu of the study’s author group, set out to test whether compact models could support on-premises deployment, where patient data never leaves the hospital, while still matching the accuracy needed for clinical information extraction.

The study analyzed 8,567 adult cervical spine CT reports collected between 2017 and 2022. Each report was labeled for acute occipital condyle and C1 through C7 fractures at the individual vertebral level, with neuroradiologist oversight. Reports from 2017, totaling 1,064, were used for finetuning, while the remaining 7,503 reports from 2018 to 2022 served as a temporal test set, allowing the researchers to check whether performance held up across successive years.

The team finetuned five small models: Gemma-3 in 1B, 4B, and 12B parameter sizes, Llama-3.2-3B, and Phi-3.5-Mini. They compared standard low-rank adaptation, a widely used efficient finetuning technique, against their new impression-regularized variant. During training, the regularized approach has the model generate multiple synthetic targeted impressions focused on acute cervical spine fractures, select the candidate most consistent with the reference standard labels, and apply an auxiliary impression loss as a regularizer. At inference, all models perform single-pass deterministic JSON extraction, producing auditable, structured output.

The results showed clear gains. Impression-regularized LoRA raised Gemma-3-4B’s macro-F1 score from 0.885 to 0.909, with macro-precision of 0.929 and macro-recall of 0.893, while also reducing variance across training runs. Improvements were largest for the smallest models and in low-data settings: Gemma-3-1B’s macro-F1 climbed from 0.646 to 0.782.

Notably, the regularized Gemma-3-4B nearly matched the zero-shot performance of the much larger Gemma-3-27B, achieving a macro-F1 of 0.909 versus 0.913. This suggests that targeted finetuning can let a model roughly six times smaller approach the extraction accuracy of a far bigger system, which matters for hospitals with limited computing resources.

Stability over time was another key finding. Across the 2018 to 2022 temporal evaluation, mean F1 increased from 0.871 to 0.893 with similar or lower variability across time windows, indicating that the regularized models did not degrade as reporting styles and patient populations shifted over the multi-year period.

The researchers also examined what the models still got wrong. Systematic error analysis showed that remaining errors were mainly associated with ambiguous report language and label-boundary cases, such as reports where the exact vertebral level of a fracture was unclear. These are inherent challenges of the source text rather than failures of the training method.

Overall, the authors conclude that impression-regularized LoRA enables small open-weight language models to extract acute cervical spine fracture labels accurately while maintaining performance across a multi-year test period and preserving auditable single-pass inference suitable for privacy-preserving on-premises deployment. As a retrospective single-center study, the findings will need validation at other institutions before broader adoption, but the approach offers a concrete path toward running capable, private, and transparent AI text extraction entirely within a hospital’s own infrastructure.

Subject of Research: Impression-regularized finetuning of small language models for extracting acute cervical spine fractures from CT radiology reports.

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

Article References: Hu, Z., Lin, H. M., Chen, Y. A., Lozano, C. S., Patel, M., Emami, A., Sejdić, E., & Colak, E. (2026). Impression-regularized finetuning enables small language models to extract cervical spine fractures from radiology reports. PLOS Digital Health, 5(10), e0001786. https://doi.org/10.1371/journal.pdig.0001786

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001786

Keywords: language models, cervical spine fractures, radiology reports, LoRA finetuning, impression regularization, on-premises deployment, CT reports, information extraction, temporal stability, Gemma-3, PLOS Digital Health, Impression-regularized