[综述笔记]Federated learning for medical image analysis: A survey

论文网址:Federated learning for medical image analysis: A survey - ScienceDirect

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. 省流版

1.1. 心得

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.2.1. Related surveys

2.2.2. Searching and analysis process

2.3. Background

2.3.1. Motivation

2.3.2. Problem formulation of federated learning

2.3.3. Typical process of federated learning

2.3.4. Types of federated learning

2.4. Federated learning for medical image analysis

2.4.1. Methods overview: A system perspective

2.4.2. Client-end learning

2.4.3. Sever-end learning

2.4.4. Client–server communication

2.5. Software platforms and tools

2.6. Medical image datasets for federated learning

2.6.1. Medical image data usage overview

2.6.2. Brain images

2.6.3. Chest/lung/heart images

2.6.4. Skin images

2.6.5. Others

2.7. Experiment

2.7.1. Experimental setup

2.7.2. Result and analysis

2.8. Discussion

2.8.1. Challenges of federated learning for medical image analysis

2.8.2. Future research directions

2.9. Conclusion

4. Reference


1. 省流版

1.1. 心得

(1)不太了解联邦学习在医学影像应用的可以读,不然不算是非常有深度的文章。可以当看个小科普。熟悉联邦学习的估计在这里面找不到新的idea和文章

2. 论文逐段精读

2.1. Abstract

        ①Conundrum in machine learning (ML): scarcity of samples

        ②Solving method: federated learning (FL)

        ③3 aspects of FL: client end, server end, and communication techniques

2.2. Introduction

        ①Privacy which limits the sharing between sites: Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR)

        ②FL in medical image field:

intractable  adj. 棘手的;很难对付(或处理)的

2.2.1. Related surveys

        ①Papers covered: from 1 January 2017 to 31 October 2023

        ②Chapters/content included: Software Platforms, Experimental Study, Future Direction and New Arisen Problems, Different Perspective and Organization

2.2.2. Searching and analysis process

        ①Introducing how the selected papers and the chapters arrangement of this work

        ②Collected/contained papers:

2.3. Background

2.3.1. Motivation

(1)Privacy protection in medical image analysis

        ①Violating privacy protection laws may result in high fines for hospitals and institutions

(2)Challenges of medical image analysis

        ①Limited sites/data

        ②Class imbalance

        ③Population statistics and distribution differences

skew  v.歪斜;偏离;歪曲;曲解;影响…的准确性;使不公允  n.斜交;扭曲;斜砌石;歪轮  adj.弯曲的;歪的;曲解的;误用的

2.3.2. Problem formulation of federated learning

        ①There are N sites with there own datasets \left \{D_{1},D_2, ... , D_N \right \}

        ②Defining the sharing model as M^* and the local model as M

2.3.3. Typical process of federated learning

(1)Client Selection and Initialization: including specific clients

(2)Local Training: advanced local training

(3)Model Upload: only update/share weights to sever

(4)Aggregation: different strategy

(5)Broadcast: synchronous update

(6)Iteration and Convergence: synchronous update

(7)Deployment: compatibility with existing hospital systems, integration challenges, and user adoption hurdles

2.3.4. Types of federated learning

(1)Horizontal federated learning

        ①Equals to homogeneous federated learning

        ②Different medical institutions hold medical imaging data of different patients

(2)Vertical federated learning

        ①Namely heterogeneous federated learning

        ②Different institutions have different types of data for the same group of patients

2.4. Federated learning for medical image analysis

2.4.1. Methods overview: A system perspective

        ①Overview:

2.4.2. Client-end learning

(1)Client end: Domain shift among clients

        ①Distribution in different sites:

        ②Domain-specific learning: fine-tune global model by data in clients

        ③Domain adaptation: align distribution differences

        ④Image harmonization: image-to-image translation model

(2)Client end: Limited data and labels

        ①Constract learning: distinguish between similar and dissimilar data

        ②Multi-task learning: data augmentation

        ③Weakly-supervised learning

        ④Knowledge distillation: utilizing small student model to represent the bigger teacher model

        ⑤Data synthesis: generative model

(3)Client end: Heterogeneous environments (computation resource & data scale)

        ①Each sites needs different epoch to train cuz their scale of samples and the device are different:

2.4.3. Sever-end learning

(1)Sever end: Weight aggregation

(2)Sever end: Domain shift among clients

(3)Sever end: Client corruption/anomaly detection

2.4.4. Client–server communication

(1)Data leakage and attack

        ①Partial Weights Sharing: only employ feature extraction in local model ()

        ②Differential Privacy: noise added

        ③Attack and Defense: generate fake images

(2)Communication efficiency

        ①Dynamic weight aggregation

2.5. Software platforms and tools

(1)PySyft

(2)OpenFL

(3)PriMIA

(4)Fed-BioMed

2.6. Medical image datasets for federated learning

2.6.1. Medical image data usage overview

        ①Using different site directly

        ②Divided one dataset to several sub dataset

2.6.2. Brain images

(1)ADNI

(2)ABIDE

(3)BraTS

(4)RSNA brain CT

(5)UK Biobank

(6)IXI

2.6.3. Chest/lung/heart images

(1)CheXpert

(2)ChestX-ray

(3)COVID-19 Chest X-ray

(4)COVIDx

(5)ACDC

(6)M&M

2.6.4. Skin images

(1)HAM10000

(2)ISIC

2.6.5. Others

(1)Eye: Kaggle Diabetic Retinopathy (Retina)

(2)Abdomen: PROMISE12

(3)Histology: TCGA

(4)Knee: fastMRI

(5)MedMNIST

2.7. Experiment

        ①Dataset: T1 weight image of ADNI

ADNI dataset containedTotal samplesADNC
ADNI 1428177229
ADNI 2360159201

        ②Atkas: AAL 90

        ③Feature: mean gray matter volumes

        ④Dataset dividing: 80% for training and 20% for testing

        ⑤Cross validation: 5 fold

 

2.7.1. Experimental setup

        ①Compared models

Traditional machine learningCrosstrain data on one client, test on other client
Singletrain and test on each client respectively
Mixtrain data from all the client, and then test
Popular FL methodsFedAVGaggregate weights
FedSGDaggregate gradients
FedProxEvery client trains its own model with an additional proximal term (the coefficient μ is set to 0.1)

        ②Figure:

2.7.2. Result and analysis

        ①Classification result:

2.8. Discussion

2.8.1. Challenges of federated learning for medical image analysis

(1)Data heterogeneity among clients: different scanner/scanning machine, patients distribution

(2)Privacy leakage/poisoning attacks

(3)Technological limitations: computing cost

(4)Long-term viability of FL-based medical image analysis

2.8.2. Future research directions

(1)Future research directions

(2)Multi-modality fusion for federated learning

(3)Model generalizability for unseen clients

(4)Weakly-supervised learning for federated learning

(5)Federated learning security: Attack and defense

(6)Blockchain and decentralization of federated learning

(7)Federated learning for medical video analysis

(8)Large-scale medical image benchmark for federated learning

(9)Model interpretability

(10)Real-world implementation and practical issues

2.9. Conclusion

        Conclusion of this work

4. Reference

Guan, H. et al. (2024) 'Federated learning for medical image analysis: A survey', Pattern Recognition, 151. doi: Redirecting

  • 8
    点赞
  • 4
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值