Research map / Vision-language and multimodal / Multimodal LLMs

EMAD: Evidence-Centric Grounded Multimodal Diagnosis for Alzheimer's Disease

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EMAD, a vision-language model that generates structured Alzheimer's diagnostic reports with sentence-to-evidence-to-anatomy grounding, trained via SFT with GTX-Distill and Executable-Rule GRPO on the AD-MultiSense dataset.

Problem

Black-box AD diagnosis from multimodal data (3D sMRI plus clinical variables) lacks explicit links from claims to clinical evidence and 3D anatomy and adherence to NIA-AA guidelines, and dense grounding annotations are expensive.

Approach

A hierarchical Sentence-Evidence-Anatomy grounding head, a teacher-to-student KL distillation (GTX-Distill) that transfers grounding from a small annotated subset to generated reports, and GRPO with executable rewards (format, NIA-AA rules, NLI-based reasoning-diagnosis consistency).

Key results

• EMAD reaches 93.33% ACC and 91.83% AUC on CN vs. CI, versus 91.28% ACC and 89.16% AUC for the best baseline M3D-LaMed.
• On three-way CN/MCI/AD classification EMAD reaches 89.4% ACC and 87.8% macro-F1, above Alifuse (84.7% ACC).
• Evidence-conditioned 3D segmentation raises hippocampus Dice from 0.78 to 0.84; GTX-Distill raises sentence-evidence R@1 from 0.57 to 0.65 over SEA without distillation.

Limitations

• Dataset is built from synthetic reports generated by DeepSeek-V3 and refined by a validator, so report quality depends on the generator.
• Zero-shot and LoRA baselines are compared under the authors' own AD-MultiSense splits; external validation is only mentioned for cross-cohort results in the appendix.

Experimental setup

Datasets: ADNI AIBL AD-MultiSense

Benchmarks: AD-MultiSense

Metrics: BLEU METEOR ROUGE BERTScore accuracy AUC sensitivity specificity macro-F1 Dice

Baselines: LLaVA-1.5-7B LLaVA-Med Med-PaLM-M M3D-LaMed Alifuse Longformer

Builds on: GRPO CLIP BLIP LoRA Segformer3D

Closest papers by meaning

• DECT: Harnessing LLM-assisted Fine-Grained Linguistic Knowledge and Label-Switched and Label-Preserved Data Generation for Diagnosis of Alzheimer’s Disease (AAAI 2025 · similarity 0.86)
• How Do Medical MLLMs Fail? A Study on Visual Grounding in Medical Images (ICLR 2026 · similarity 0.86)
• Beyond Classification Accuracy: Neural-MedBench and the Need for Deeper Reasoning Benchmarks (ICLR 2026 · similarity 0.86)
• RAD: Towards Trustworthy Retrieval-Augmented Multi-modal Clinical Diagnosis (NeurIPS 2025 · similarity 0.86)
• OmniBrainBench: A Comprehensive Multimodal Benchmark for Brain Imaging Analysis Across Multi-stage Clinical Tasks (CVPR 2026 · similarity 0.86)
• Toward a Vision-Language Foundation Model for Medical Data: Multimodal Dataset and Benchmarks for Vietnamese PET/CT Report Generation (NeurIPS 2025 · similarity 0.86)