MAVeRiC-AD: Mixture-of-experts Agentic Vision-Language Ensemble for Robust MRI Classification of Alzheimer's Disease

Abstract

Robust classification of Alzheimers disease (AD) from structural T1-weighted MRI (T1w) images remains an unmet clinical need, especially when data is acquired at multiple sites that differ in scanning protocols and population demographics. In this paper, we present MAVeRiC-AD (Mixture-of-experts Agent-guided Vision-Language Ensemble for Robust Imaging-based classification of Alzheimers Disease), an agentic framework that dynamically utilizes the optimal inferencing tool for radiological queries to provide relevant answers to the user. In our framework, we tested three specialized models encoded as callable tools: (1) CNN-AD, a 3D DenseNet trained on T1w intensities only; (2) MOE-VLM, a vision-language model that jointly models the T1w with subject-specific demographics (age, sex, site) via a mixture-of-experts (MoE) projection head; (3) Retrieval engine, a similarity-search module that contextualizes a patient against others from the site and reports % prevalence of AD. A light-weight agent analyzes the user request and then routes the input (image, or image + text) to the appropriate tool, aggregates responses and returns the tool response augmented with its confidence derived from conformal prediction. Experiments were conducted using T1w images from the ADNI (N=4,098) and OASIS-3 (N=600) datasets. Single-site training baselines achieved ROC-AUC = 0.79 (CNN) and 0.82 (VLM) on ADNI. When trained jointly on both sites, MOE-VLM surpassed both image-only and standard vision-language models with ROC-AUC = 0.90 on ADNI and 0.81 on OASIS. MAVeRiC-AD demonstrates that agentic orchestration of complementary expert deep models, coupled with explicit demographic conditioning for multi-site data can improve robustness and interpretability of AD image analysis pipelines and serves as a blueprint for scalable, trustworthy clinical AI assistants.