The Radiology Sparring Partner: A PoC Study of Multi-Agentic Workflows for Clinical Decision Support and Teaching
- 1Cognitive-Systems, Otto-von-Guericke University Magdeburg, Universitätsplatz 2,Magdeburg, Germany
- 2Department of Radiology, Otto-von-Guericke University Magdeburg, Leipziger Str. 44,Magdeburg, Germany
- 3Department of Psychosomatic Medicine and Psychotherapy, Otto-von-Guericke University Magdeburg, Leipziger Str. 44,Magdeburg, Germany
Abstract
Integrating Large Language Models (LLMs) into radiology is often hindered by the inability of monolithic architectures to balance clinical reasoning with pedagogical safety. This study evaluates MedGemma-27B-IT across two specialized agentic frameworks: a clinical "Radiology Sparring Partner" and a "Socratic Teacher". For clinical decision support, we compared three architectures: Raw (base model output, no prompt), Minimal Prompted (single-shot prompt), and Multi-Agentic, using a 4-stage pipeline for clinical synthesis. Separately, the educational track employs a dual-agent architecture for progressive educational scaffolding. A pooled evaluation of 120 interaction transcripts (10 per architecture with three independent LLM-as-a-judge runs) was finalized using an expert-anchored combined mean approach, anchored via Bühlmann credibility calibration against radiologist-authored references (5 per architecture). Results demonstrate that agentic orchestration outperforms monolithic baselines in high-level reasoning. Most notably, Pathophysiological Logic scores rose from 3.20 (Raw) to 4.14 (Agentic) on a 5-point Likert scale, with a simultaneous reduction in standard deviation from 1.23 to 0.70. While this Proof of Concept (PoC) highlights the advantages of modularity, results remain subject to evaluator formulation bias and a measurable "latency tax" inherent in sequential agent hand-offs. Future work will focus on longitudinal clinical validation and optimization of cached memory architectures to refine reasoning depth while mitigating latency.
Related
- RadKev: An Open-Weight Decision Model for Radiology
- Can Jev Judge Radiology Reports? Evaluating a System One Model for Clinical Factuality
- Advances in Multimodal Artificial Intelligence in Radiology: Data Integration, Foundation Models, and Clinical Applications-A Narrative Review.
- Effects of an LLM-based intelligent adaptive teaching program on clinical risk perception among newly employed nurses: a quasi-experimental study
- A prebill control plane for record integrity in inpatient artificial intelligence-enabled clinical documentation workflows.
BibTeX
@article{jonas2026radiology,
title = {{The Radiology Sparring Partner: A PoC Study of Multi-Agentic Workflows for Clinical Decision Support and Teaching}},
author = {Schewior Jonas and Ashfaq Anwerdeen Mohamed and Ertürk Birim and Großer Oliver and Siegert Ingo and Pech Maciej and Wendemuth Andreas},
journal = {Current Directions in Biomedical Engineering},
year = {2026},
month = oct,
doi = {10.1515/cdbme-2026-0216},
url = {https://doi.org/10.1515/cdbme-2026-0216},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - The Radiology Sparring Partner: A PoC Study of Multi-Agentic Workflows for Clinical Decision Support and Teaching
AU - Schewior Jonas
AU - Ashfaq Anwerdeen Mohamed
AU - Ertürk Birim
AU - Großer Oliver
AU - Siegert Ingo
AU - Pech Maciej
AU - Wendemuth Andreas
PY - 2026
DA - 2026/10/26
JO - Current Directions in Biomedical Engineering
DO - 10.1515/cdbme-2026-0216
UR - https://doi.org/10.1515/cdbme-2026-0216
AB - Integrating Large Language Models (LLMs) into radiology is often hindered by the inability of monolithic architectures to balance clinical reasoning with pedagogical safety. This study evaluates MedGemma-27B-IT across two specialized agentic frameworks: a clinical "Radiology Sparring Partner" and a "Socratic Teacher". For clinical decision support, we compared three architectures: Raw (base model output, no prompt), Minimal Prompted (single-shot prompt), and Multi-Agentic, using a 4-stage pipeline for clinical synthesis. Separately, the educational track employs a dual-agent architecture for progressive educational scaffolding. A pooled evaluation of 120 interaction transcripts (10 per architecture with three independent LLM-as-a-judge runs) was finalized using an expert-anchored combined mean approach, anchored via Bühlmann credibility calibration against radiologist-authored references (5 per architecture). Results demonstrate that agentic orchestration outperforms monolithic baselines in high-level reasoning. Most notably, Pathophysiological Logic scores rose from 3.20 (Raw) to 4.14 (Agentic) on a 5-point Likert scale, with a simultaneous reduction in standard deviation from 1.23 to 0.70. While this Proof of Concept (PoC) highlights the advantages of modularity, results remain subject to evaluator formulation bias and a measurable "latency tax" inherent in sequential agent hand-offs. Future work will focus on longitudinal clinical validation and optimization of cached memory architectures to refine reasoning depth while mitigating latency.
SN - 2364-5504
ER -