Real-time structuring of the clinical information section for dopamine transporter SPECT reporting with a small language model pipeline: development and external validation

Background Accurate interpretation of dopamine transporter (DAT)-SPECT depends on clinical context, including symptoms, medication, comorbidities, and structural imaging findings. In routine practice, however, this information is often documented as heterogeneous free text, which may hinder consistent consideration of clinically relevant parameters and limit guideline-compliant reporting. We therefore developed and evaluated a locally deployable small language model (SLM)-based pipeline to structure and standardize the clinical information section for DAT-SPECT reporting. Two retrospective datasets were used. The in-house dataset included clinical information sections from 2,990 DAT-SPECT examinations and was split into training ( n = 2,700) and internal test data ( n = 290). The external validation dataset comprised 595 clinical information sections from a second institution. A structured dictionary (466 entries across 12 categories) was created by two experienced DAT-SPECT readers and used to manually annotate the training and internal test samples. A Qwen3 1.7B model was fine-tuned by supervised learning to reproduce expert annotations. A rule-based postprocessing pipeline converted model outputs into report-ready structured text, including removal of redundant entries, automated suggestion of scan indication, consolidation of symptom lateralization and time course, reminders for missing guideline-relevant elements, and medication interaction checks. Internal test performance was assessed by expert comparison of SLM output with manual annotation. External performance was assessed by counting the manual corrections required in the postprocessed output under strict criteria. Results In the internal test dataset, the postprocessed SLM output provided the same information as the expert annotation in 196/290 cases (67.6%), including exact matches in 147/290 (50.7%). Interpretation differences due to free-text ambiguity occurred in 13/290 cases (4.5%), and SLM errors in 41/290 cases (14.1%). In 40/290 cases (13.8%), the SLM output was judged superior to the manual annotation. Among the error cases, 36/41 (87.8%) involved only a single erroneous or missing item. In the external dataset, 446/595 postprocessed outputs (75.0%) required no manual correction before inclusion in the report. Among the remaining 149/595 cases (25.0%), correction of only a single item was required in 111 cases (18.7% of all cases). The mean number of corrections was 0.34 items per case, corresponding to 6.0% of all structured items. Conclusions A locally deployable SLM combined with rule-based postprocessing can transform free-text clinical information for DAT-SPECT into standardized, report-ready documentation with good practical usability in an external dataset. End-to-end processing times of typically less than one second on the tested hardware support the feasibility of real-time local deployment. Prospective studies are required to determine its effects on clinical workflow and reporting quality.

Authors

Publication Details

Journal
EJNMMI Research
Published
2026-11-19
DOI
https://doi.org/10.1186/s13550-026-01542-5
Primary Topic
Artificial Intelligence in Healthcare and Education
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Real-time structuring of the clinical information section for dopamine transporter SPECT reporting with a small language model pipeline: development and external validation

Rickmer Braren, Ivayla Ilieva Apostolova, Susanne Klutmann, Ralph Buchert et al.
1 links
EJNMMI Research
Artificial Intelligence in Healthcare and Education
article

Real-time structuring of the clinical information section for dopamine transporter SPECT reporting with a small language model pipeline: development and external validation

Rickmer Braren, Ivayla Ilieva Apostolova, Susanne Klutmann, Ralph Buchert, Thomas Buddenkotte, Philipp T. Meyer
article en

Abstract

Background Accurate interpretation of dopamine transporter (DAT)-SPECT depends on clinical context, including symptoms, medication, comorbidities, and structural imaging findings. In routine practice, however, this information is often documented as heterogeneous free text, which may hinder consistent consideration of clinically relevant parameters and limit guideline-compliant reporting. We therefore developed and evaluated a locally deployable small language model (SLM)-based pipeline to structure and standardize the clinical information section for DAT-SPECT reporting. Two retrospective datasets were used. The in-house dataset included clinical information sections from 2,990 DAT-SPECT examinations and was split into training ( n = 2,700) and internal test data ( n = 290). The external validation dataset comprised 595 clinical information sections from a second institution. A structured dictionary (466 entries across 12 categories) was created by two experienced DAT-SPECT readers and used to manually annotate the training and internal test samples. A Qwen3 1.7B model was fine-tuned by supervised learning to reproduce expert annotations. A rule-based postprocessing pipeline converted model outputs into report-ready structured text, including removal of redundant entries, automated suggestion of scan indication, consolidation of symptom lateralization and time course, reminders for missing guideline-relevant elements, and medication interaction checks. Internal test performance was assessed by expert comparison of SLM output with manual annotation. External performance was assessed by counting the manual corrections required in the postprocessed output under strict criteria. Results In the internal test dataset, the postprocessed SLM output provided the same information as the expert annotation in 196/290 cases (67.6%), including exact matches in 147/290 (50.7%). Interpretation differences due to free-text ambiguity occurred in 13/290 cases (4.5%), and SLM errors in 41/290 cases (14.1%). In 40/290 cases (13.8%), the SLM output was judged superior to the manual annotation. Among the error cases, 36/41 (87.8%) involved only a single erroneous or missing item. In the external dataset, 446/595 postprocessed outputs (75.0%) required no manual correction before inclusion in the report. Among the remaining 149/595 cases (25.0%), correction of only a single item was required in 111 cases (18.7% of all cases). The mean number of corrections was 0.34 items per case, corresponding to 6.0% of all structured items. Conclusions A locally deployable SLM combined with rule-based postprocessing can transform free-text clinical information for DAT-SPECT into standardized, report-ready documentation with good practical usability in an external dataset. End-to-end processing times of typically less than one second on the tested hardware support the feasibility of real-time local deployment. Prospective studies are required to determine its effects on clinical workflow and reporting quality.

EJNMMI ResearchVol. 16(1)
Good health and well-being
Artificial Intelligence in Healthcare and Education
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon