论文网址：Large Language Models Improve Alzheimer's Disease Diagnosis Using Multi-Modality Data | IEEE Conference Publication | IEEE Xplore

英文是纯手打的！论文原文的summarizing and paraphrasing。可能会出现难以避免的拼写错误和语法错误，若有发现欢迎评论指正！文章偏向于笔记，谨慎食用

目录

1. 心得

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.3. Materials and Methods

2.3.1. Materials

Clinical Data

Neuropsychological Data

Biospecimen Results

Genetic Data

Medical History

Family History

2.3.2. Embedding of Non-Image Data

2.3.3. Prompt Engineering

2.3.4. Modality Alignment

2.3.5. Model

2.4. Experiments and Results

2.4.1. Data Description

2.4.2. Experimental Settings

2.4.3. Adni-2 Classification Results

2.4.4. Effectiveness of Llm

2.4.5. Ablation Experiments

2.5. Conclusions

1. 心得

（1）好短，双栏六页

（2）虽然这篇论文不是发在很好的会议上，但我也想说作者把图像特征直接输入了GPT。因为现在没什么fmri和mri的大模型

2. 论文逐段精读

2.1. Abstract

        ①LLM can be used to extract the non-image information of patients

2.2. Introduction

        ①Limitations: in quite a lot ablation studies, phenotypic data represents few influence to final results, which means limit development of non-image data

2.3. Materials and Methods

2.3.1. Materials

        ①Included image data: MRI and PET:

        ②Included non-image data:

Clinical Data	Diagnosis and Symptoms Checklist
	Diagnostic Summary
Neuropsychological Data	ADSP-PHC Composite Cognitive Scores
	Functional Activities Questionnaire (FAQ)
	Mini-Mental State Examination (MMSE)
	Montreal Cognitive Assessment (MoCA)
Biospecimen Results	APOE - Results
	Fujirebio Beta-Amyloid Ratio
	CSF - Local Lab Results
Genetic Data	Desikan Lab Polygenic Hazard Score (PHS)
Medical History	Adverse Events/Hospitalizations
	Documentation of Baseline Symptoms Log
	Initial Health Assessment
Family History	NO SCALE CUZ OF HIGH RELEVANT

2.3.2. Embedding of Non-Image Data

        ①Performance of sequential model:

2.3.3. Prompt Engineering

        ①Working pipeline:

2.3.4. Modality Alignment

        ①Employ cross attention mechanism on image modality and non-image modality fusion

2.3.5. Model

        ①Framework of proposed model:

        ②Backbone network: ConvNeXt-tiny (Vision Encoder) with 512 dimension of feature output

2.4. Experiments and Results

2.4.1. Data Description

        ①Classes: early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI), normal control (NC), and AD

        ②Samples in ADNI-2: 515, 80% for training and 20% for testing

        ③Pre-training: adopt 103 subjects in ADNI 1

2.4.2. Experimental Settings

        ①Lists metrics and category setting (3 classes and 4 classes tasks)

2.4.3. Adni-2 Classification Results

        ①Performance comparison with other models:

        ②Performance of proposed model in different classification tasks:

2.4.4. Effectiveness of Llm

        ①Module ablation:

2.4.5. Ablation Experiments

        ①Ablation studies:

2.5. Conclusions

        ~