AutoProtocol: Agent-Driven Automation of CT and MRI Protocoling in Radiology

Philipp Wesp, Louisa Fay, Faidra Patsatzi, Sergios Gatidis

Keywords: Imaging Exam Protocoling, In-Context Learning, LLM-agents, Radiology

TL;DR: AutoProtocol is an agent-based training-free approach to generate protocols for CT and MR based on in-context learning, including patient history and similar cases from a 4.4M-exam database. On 500 cases, it achieves over 73% Hits@1.

Registration Requirement: Yes

Abstract:

In radiology, imaging exam protocoling is a high-volume, expert-driven task that increasingly limits clinical throughput as imaging demand grows.
We introduce AutoProtocol, an agentic pipeline that frames protocol generation as a training-free, retrieval-augmented problem.
Given a patient’s history, the system retrieves similar prior cases from large-scale clinical data and provides them as context to an LLM-based agent, enabling patient-specific protocol generation without task-specific fine-tuning.
Using a dataset of 4.4M exams, AutoProtocol achieves strong performance on 500 evaluation cases, reaching 0.84 mean reciprocal rank (MRR) and 76.1\% Hits@1 under LLM-based evaluation, and 0.81 MRR / 73.6\% Hits@1 using embedding-based similarity (vs. 0.10 MRR / 10\% Hits@1 baseline).

Reproducibility: AutoProtocol frames radiology protocol selection as a retrieval-augmented in-context learning problem: patient history and similar historical cases from a 4.4M-exam database are retrieved at inference time and used to condition an off-the-shelf LLM, achieving 76.1% Hits@1 on 500 CT/MRI cases with no task-specific training.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 111