论文网址：Deep learning for brain disorder diagnosis based on fMRI images - 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. 心得

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.3. The overview of deep learning methods

2.3.1. Artificial intelligence, machine learning and deep learning

2.3.2. Brief introduction to deep learning techniques in fMRI analysis

2.4. Deep learning in brain disorder diagnosis

2.4.1. Functional connectivity model based approaches

2.4.2. 2D/3D image processing perspective

2.4.3. fMRI images as a time series

2.4.4. Joint spatial and temporal feature exploration

2.4.5. Other deep learning models and training techniques

2.4.6. Summary

2.5. Challenges and future outlook

2.5.1. Discussions

2.5.2. Future outlook

3. Reference

---

1. 心得

（1）少儿科普是吧？不推荐已经很熟悉这个领域的人看

（2）我怎么才刚开始就要看不下去了

（3）仅用图表示最基础的结构（如CNN或AE），但用过量的文字讲述一万个模型，是一种，让人，极为没有耐心的写作方式

（4）有一种站在未来看过去的感觉...但是吧...有些过去不是非要看的...

2. 论文逐段精读

2.1. Abstract

        ①fMRI can be regarded as image, time series and image series

2.2. Introduction

        ①Introducing fMRI itself→applications→pros and cons→compared with other imaging methods→introducing ML and AI

2.3. The overview of deep learning methods

2.3.1. Artificial intelligence, machine learning and deep learning

（1）Artificial intelligence

        ①介绍了...AI...

        ②The relationship between AI, ML and DL:

（2）Machine learning

        ①Categories of ML: supervised learning, unsupervised Learning, semi-supervised learning and reinforcement learning

（3）Deep learning

        ①...有出色的性能？？受益于计算能力？？fine

2.3.2. Brief introduction to deep learning techniques in fMRI analysis

        ①Common use DNN: CNN and RNN

        ②Workflow of deep learning:

（1）Convolutional neural networks

        ①介绍CNN？？？？卷积？？！激活函数？？

        ②图像分类的简化CNN网络？？？图？和上一张图有什么不一样？？

（2）Recurrent neural network

        ①Famous RNN: LSTM and GRU

（3）Auto encoder and decoder

        ①An example AE:

2.4. Deep learning in brain disorder diagnosis

2.4.1. Functional connectivity model based approaches

（1）Functional connectivity model construction

        ①Static and dynamic FC construction methods:

（2）FC based DL methods

        ①ML pipeline:

（3）Direct use of functional connectivity measures

        ①Directly classify FC matrix

        ②疯狂地用文字介绍了一堆方法

（4）End-to-end model for disease classification

        ①继续介绍一堆模型

（5）Connectivity matrices as an analogy to 2D images

        ①...Some researchers consider fMRI as 2D images and Conv it

（6）Comment on functional connectivity based methods

        ①There are some noises in raw fMRI data

2.4.2. 2D/3D image processing perspective

（1）4D fMRI to 2D images conversion

        ①列举了一堆4D转2D然后卷积的办法，从引用就能看出这些办法其实很老了，全在2020之前，而且这本来就有那么点不合理（虽然这篇是22写的啦...所以其实对于现在的我们参考性不是很强）

（2）3D neural network

        ①扒拉了一堆模型

（3）Challenges and opportunities

        ①For limited samples, the prior kownledge is feasible

2.4.3. fMRI images as a time series

        ①继续列

2.4.4. Joint spatial and temporal feature exploration

        ①Lists models with spatial and temporal methods

2.4.5. Other deep learning models and training techniques

（1）Graph CNN and its applications

（2）Generative models

（3）Transfer learning

2.4.6. Summary

        ①作者觉得以后深度学习会在临床诊断中大放异彩

2.5. Challenges and future outlook

2.5.1. Discussions

        ①Non of a DL model can contain all the tasks of disease diagnosis（确实，作者当时写的时候大模型还没有风靡，可能确实觉得，一个模型很难顾及到所有方面）

        ②Challenges: a) costs, b) interpretability

2.5.2. Future outlook

        ①“尽管近年来使用 fMRI 图像在脑部疾病诊断方面取得了巨大成功，但距离临床诊断要求还很远”，没错其实...2024年仍然还有一些距离，还需要一点重大突破...

        ②Researchers should mix all the data together, including electronic medical record, EEG, structural MRI image etc...确实，现在多模态还发展得不错

3. Reference

Yin, W., Li, L., & Wu, F. (2022) 'Deep learning for brain disorder diagnosis based on fMRI images', Neurocomputing, 469: 332-345. doi: https://doi.org/10.1016/j.neucom.2020.05.113