Abstract
We propose a fully automated pipeline for brain MRI volumetry report generation and hallucination detection using lightweight, locally deployable LLMs. Starting from quantitative volumetric outputs of a brain segmentation tool, a generator LLM produces six-section clinical reports under eleven progressively enriched prompt configurations. A separate out-of-family judge LLM audits each report against a formal clinical rule set, classifying violations into clinically dangerous (Type A) and format compliance (Type B) errors. Judge reliability is validated on adversarially corrupted reports against a frontier-model gold standard. Our ablation study across four generator models (Llama-3, Ministral-3, each at 3B/8B) shows that clinical hallucination rates drop from 54-96% at baseline to 0-4% with appropriate prompting. Anchoring-based verification achieves acceptable clinical hallucination (95% CI: 0-7%) in two configurations. Chain-of-thought reasoning, beneficial for 8B models, is systematically detrimental for 3B generators. The report generation runs locally on commodity hardware under 40 seconds, thereby satisfying clinical data protection requirements without requiring cloud connectivity.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ELAMI_001.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/ELAMI_001_supp.pdf
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{CouPie_RuleCompliant_MICCAISAT2026,
author = { Coupé, Pierrick AND Zamai, Andrew AND Mansencal, Boris AND Planche, Vincent AND Tourdias, Thomas AND Morandat, Floréal AND Simon, Laurent AND Fijalkow, Nathanaël AND Manjón, José},
title = { { Rule-Compliant Brain MRI Volumetry Report Generation with Locally Deployable LLMs } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}