Evidence-Supported Automated Impressions for Alzheimer’s Disease Detection from Brain MRI: A Feasibility Study

Devesh Singh, Dhanush H. Babu, Mathias Manzke, Marc-André Weber, Majid Ramedani, Arijana Bohr, Emmanuelle Salin, Bjoern Eskofier, Martin Dyrba

Keywords: LLM, Radiological Impressions, Dementia, MRI, RAG, Hallucination

TL;DR: We present an end-to-end pipeline for automated MRI-based dementia impression generation and a falsifiable RAG hallucination taxonomy, demonstrating that claim-level inaccuracies persist despite evidence grounding.

Registration Requirement: Yes

Abstract: Convolutional Neural Networks (CNNs) are the standard models for neuroimaging analysis, but their opacity hinders clinical adoption. While Large Language Models (LLMs) offer a potential solution to translate CNN outputs as human-readable impressions, their reliability remains questionable. In the context of Alzheimer's Disease (AD) detection, we introduce a framework that computes CNN-based brain morphology scores and leverages a rule module for summarization. Using an LLM, conditioned on diagnostic guidelines via Retrieval-Augmented Generation (RAG), we generate explanatory justifications of pathology detected. To assess the impact of hallucinations, we propose a taxonomy that considers generated justifications as falsifiable claims. Manual evaluation on 30 reports shows that hallucinations remain substantial. In the pathological cases, from 34–50\% of claims were incorrect. Our feasibility study shows that integrating rule-based guardrails with RAG improves auditability but fails to sufficiently mitigate hallucinations.

Reproducibility: https://github.com/DhanushBabu18/MRIs-to-Radiological-Dementia-Reports/

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 31