Semantic retrieval-augmented translation for radiology reports using small language models: Algorithm development and evaluation
GMS Medizinische Informatik, Biometrie und Epidemiologie · 22.09.2026 · 22 (Doc15) · Doc15
DOI: 10.3205/mibe000313 ↗Kurz gefasst
Fachartikel in der Zeitschrift GMS Medizinische Informatik, Biometrie und Epidemiologie, erschienen am 22.09.2026. Verfasst von Reichenpfader und 4 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Introduction
Accurate machine translation (MT) of clinical text is essential for multilingual patient care, cross-border data sharing, and research reproducibility, yet the use of commercial APIs for such tasks raises substantial privacy concerns. We present Semantic Retrieval-Augmented Translation (S-RAT), a lightweight, ontology-guided framework that integrates the Unified Medical Language System (UMLS) and SNOMED CT concept retrieval into large language model prompts to improve medical terminology fidelity without fine-tuning or external APIs.
Methods
Using 500 English radiology reports from MIMIC-CXR, we evaluated multiple S-RAT variants with open-source language models (gemma3, med42-v2, deepseek-r1, and gpt-oss) as well as a commercial frontier model (Gemini 3 Pro). We compared reference-free quality estimation metrics (COMET-src and GEMBA-DA-noref) with human annotation. Furthermore, we assessed the translation coverage of the German SNOMED CT edition for the specific use case of radiology report translation.
Results
S-RAT prompting did not consistently outperform standard prompting, and COMET-src correlated only weakly (ρ≈0.14, p≈0.05) with human adequacy judgments, while GEMBA-DA-noref showed no meaningful correlation. The German SNOMED CT edition covered only 32.6% of extracted concepts.
Discussion
These findings highlight the current limitations of multilingual ontology resources and the application of general-domain MT metrics in clinical contexts. S-RAT demonstrates the feasibility of privacy-preserving, locally deployable, and ontology-augmented translation and provides a foundation for future work on domain-specific evaluation, ontology enrichment, and local deployment of language models in healthcare settings.
Schlagwörter
Zitieren
APA
Reichenpfader, Daniel, Dennstä, dt, & Fabio (2026). Semantic retrieval-augmented translation for radiology reports using small language models: Algorithm development and evaluation. GMS Medizinische Informatik, Biometrie und Epidemiologie, 22(Doc15), Doc15. https://doi.org/10.3205/mibe000313
Vancouver
Reichenpfader, Daniel, Dennstä, dt, Fabio. Semantic retrieval-augmented translation for radiology reports using small language models: Algorithm development and evaluation. GMS Medizinische Informatik, Biometrie und Epidemiologie. 2026;22(Doc15):Doc15. doi:10.3205/mibe000313
BibTeX
@article{Reichenpfader2026,
author = {Reichenpfader and Daniel and Dennstä and dt and Fabio},
title = {Semantic retrieval-augmented translation for radiology reports using small language models: Algorithm development and evaluation},
journal = {GMS Medizinische Informatik, Biometrie und Epidemiologie},
year = {2026},
volume = {22},
number = {Doc15},
pages = {Doc15},
doi = {10.3205/mibe000313}
}
EUVIMED