A locally deployable artificial intelligence pipeline powered by a small language model can standardize unstructured clinical records for dopamine transporter single-photon emission computed tomography reporting in under one second, according to research published in EJNMMI Research. Evaluated across retrospective patient datasets from two independent medical institutions, the system converted complex free-text medical notes into structured, report-ready documentation. In an external validation cohort, three-quarters of the generated clinical summaries required zero manual corrections before being incorporated directly into final medical reports, demonstrating substantial practical usability for diagnostic workflows.
However, accurate visual and quantitative interpretation of these nuclear medicine scans depends heavily on comprehensive clinical context, such as current patient symptoms, active medications, existing comorbidities, and prior structural imaging findings. In routine clinical practice, referring healthcare providers typically document this essential background as unstructured, heterogeneous free text. This lack of standardization frequently obscures vital clinical details, complicates systematic evaluation, and hinders consistent adherence to established nuclear medicine reporting guidelines.
The development process relied on two distinct retrospective collections of clinical records. The primary in-house dataset comprised clinical information sections from 2,990 DAT-SPECT examinations performed at the primary center. The researchers divided this collection into a training partition of 2,700 cases to teach the model and an internal test set of 290 cases reserved for performance evaluation. An additional external validation dataset containing 595 clinical sections was gathered from a second independent institution to verify model generalizability.
Establishing a reliable ground truth for training required extensive manual clinical curation. Two experienced DAT-SPECT readers constructed a standardized medical dictionary containing 466 distinct entries organized across 12 specific clinical categories. The specialists used this structured schema to manually annotate the in-house training and internal test records.
Raw model predictions were subsequently processed by an automated rule-based postprocessing pipeline to transform preliminary extractions into polished, clinically reliable documentation. This postprocessing phase removed redundant text entries and automatically proposed the primary clinical indication for each requested scan.
Internal testing on the 290 in-house evaluation cases demonstrated close alignment between automated extractions and expert curation. When researchers assessed the postprocessed outputs, the model provided identical clinical information to the expert annotations in 196 out of 290 cases, representing 67.6 percent of the test cohort, including exact textual matches in 147 cases, or 50.7 percent. Differences arising from genuine ambiguity in the source free text occurred in 13 cases, or 4.5 percent, while distinct model errors appeared in 41 cases, or 14.1 percent. In 40 cases, or 13.8 percent of the cohort, reviewers judged the model output superior to the original manual annotation. Among the cases with errors, 36 out of 41, or 87.8 percent, involved merely a single erroneous or omitted entry.
External testing on the 595 cases from the second clinical institution confirmed that the automated pipeline maintained consistent performance across differing documentation habits. When evaluated under strict criteria, 446 out of 595 postprocessed summaries, or 75.0 percent, required no manual correction whatsoever before report inclusion. Among the remaining 149 cases where adjustments were necessary, 111 cases required the correction of only a single item, accounting for 18.7 percent of the entire external cohort. Across all 595 external cases, the mean correction rate was 0.34 items per patient encounter, which corresponded to 6.0 percent of all structured data fields generated by the system.
Measured end-to-end processing times consistently remained below one second on the tested computing hardware, indicating that the pipeline can operate in real time within routine clinical radiology environments. The authors concluded that while retrospective performance shows strong feasibility, future prospective clinical trials are necessary to directly quantify the technology’s influence on daily reporting efficiency, workflow integration, and overall diagnostic reporting quality.
Subject of Research: Automated structuring of clinical information for dopamine transporter SPECT reporting using small language models
Article Title: Real-time structuring of the clinical information section for dopamine transporter SPECT reporting with a small language model pipeline: development and external validation
Article References: Buddenkotte, T., Apostolova, I., Klutmann, S., Braren, R., Meyer, P. T., & Buchert, R. (2026). Real-time structuring of the clinical information section for dopamine transporter SPECT reporting with a small language model pipeline: development and external validation. EJNMMI Research, 16(1), Article 164. https://doi.org/10.1186/s13550-026-01542-5
Image Credits: AI Generated
DOI: 10.1186/s13550-026-01542-5
Keywords: dopamine transporter SPECT, small language models, clinical information structuring, Qwen3 1.7B, medical report generation, rule-based postprocessing, external validation, supervised fine-tuning, medical text annotation, symptom lateralization, medication interaction checks, neuroimaging informatics, clinical free-text processing, diagnostic guideline compliance, retrospective dataset validation, real-time clinical deployment
News Source: Ophelia Keating. (November 19, 2026). Small Language Model Pipeline Accurately Structures Clinical Notes for Dopamine SPECT. Scienmag.



