[论文精读]FCNet: A Convolutional Neural Network for Calculating Functional Connectivity from Functional

夏莉莉iy

已于 2024-03-05 22:38:27 修改

阅读量954

点赞数 20

分类专栏：论文精读文章标签：深度学习机器学习人工智能计算机视觉学习神经网络 python

于 2024-03-04 23:51:12 首次发布

本文链接：https://blog.csdn.net/sherlily/article/details/136459360

版权

论文精读专栏收录该内容

98 篇文章 10 订阅

订阅专栏

本文介绍了一种名为FCNet的卷积神经网络，用于直接从功能磁共振成像时间序列中提取功能性连接。研究者提出了一种新颖的方法，通过训练生成FC并结合特征选择和SVM进行ADHD分类，结果显示其在ADHD分类竞赛中表现出色。

摘要由CSDN通过智能技术生成

论文全名：FCNet: A Convolutional Neural Network for Calculating Functional Connectivity from Functional MRI

论文网址：FCNet: A Convolutional Neural Network for Calculating Functional Connectivity from Functional MRI | SpringerLink

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

2.3.1. Data and Preprocessing

2.3.2. Functional Connectivity Through FCNet

2.3.3. Feature Selection and Classification

2.4. Experiments and Results

2.5. Conclusion

3. Reference List

1. 省流版

1.1. 心得

（1）FC时间上相关的啊，意思空间相关不能叫FC咯？确实，想想也不是没道理，fmri本来就是功能影像（时间），smri才是结构（空间）

（2）很久没看到语言上这么奇怪的论文了，怎么感觉和我不是一个世界的人

（3）以前古早的可视化确实做的不是那么好，可以添加颜色捏

1.2. 论文总结图

2. 论文逐段精读

2.1. Abstract

①Traditional correlation and distance measures are essential for machine learning in fMRI technique. However, these method may not caputure precise time series feature

②Thus, the authors put forward FCNet, which extract functional connectivity (FC) from time signals directly

③They aim to classify ADHD and HC

2.2. Introduction

①FC presents the pair-wise temporal correlation of each brain regions

②⭐The approaches of FC constructions are correlation, clustering, graph based etc...

③The overall framework of FCNet:

(a) Train FCNet by data, which generated by generator;

(b) Generating FC by FCNet, selecting features and connecting them with phenotypic data. Then training SVM

（我天这段话..怎么...有那么点...绕口...）

2.3. Method

2.3.1. Data and Preprocessing

①Dataset: ADHD-200

②Sites: NeuroImage (NI), New York University Medical Center (NYU) and Peking University (Peking)

③Atlas: automated anatomical labelling atlas (AAL) 90

2.3.2. Functional Connectivity Through FCNet

①FCNet is a deep-network architecture

②The architecture of FCNet:

（1）The Feature Extractor Network

①Activate function: Leaky Rectified Linear Unit (ReLU), cuz it converges faster than ReLU

②Length of time-series signal: 172

③Pool length in pooling layer: 2

④Kernal size: 3

⑤Filters of C 1-5: 32, 64, 96, 64, 64

⑥The size of the last fully connected layer (purple blocks): 32

（2）The Similarity Measure Network

①Extracting correlation by neural network

②The essence of building relevance is to utilize similarity

③The last layer contains a Softmax

（3）Coupled Architecture with Shared Parameters

①They share the parameters in the coupled structure（呃呃为什么要写得这么抽象，所以意思是图上两根蓝线是一个人的两个ROI吗？）

（4）Data Generator for Training FCNet

①⭐They need similar regions, which means functionally connected and dissimilar regions, which means not connected functionally. Labelling similar one with 1, dissimilar one with 0（救命...这是什么）

②Labelling 1 to ROIs with the same communities and 0 with dissimilar communities

③The specific algorithm:

（5）Training of FCNet

①Loss function: cross-entropy

$L_{fc}=-\frac{1}{n}\sum_{1}^{n}[y_{i}log(\hat{y}_{i})+(1-y_{i})log(1-\hat{y}_{i})]$

where $n$ denotes the number of training samples, $y_i \in \left \{ 0,1 \right \}$ , $\hat{y}_i$ denotes the predicted result

2.3.3. Feature Selection and Classification

①Feature selection: Elastic Net (EN) based

②EN combines $L_1$ and $L_2$ penalty that $L_1$ brings variable selection and continuous shrinkage and $L_2$ encourages grouped selection of features

③Defining $\boldsymbol{y}$ as the vector of subjects $y_{i}\epsilon(l_{1},l_{2},...l_{n})$ and $\boldsymbol{X}=\{FC_{1},FC_{2},...FC_{n}\}$ , the cost function will be:

$L_{en}(\lambda_1,\lambda_2,\beta)=(||\boldsymbol{y}-\boldsymbol{X}\boldsymbol{\beta}||)^2+\lambda_1(||\boldsymbol{\beta}||)_1+\lambda_2||\boldsymbol{\beta}||^2$

where $\lambda_1,\lambda_2$ are weights;

$\beta$ needs to be calculated

2.4. Experiments and Results

①Classes: HC and ADHD (combined with ADHD combined, ADHD hyperactive-impulsive and ADHD inattentive)

②The model proposed by the authors outperforms all other models in the competition. And for datasets from different sites, the parameters in their frameworks remain unchanged.

③Comparisons in competition:

④Validation of phenotype data usage:

⑤Visualization of FC comparison between HC and ADHD on Peking site:

2.5. Conclusion

阐述了一下它究竟有多么复杂

3. Reference List

Riaz, A. et al. (2017) 'FCNet: A Convolutional Neural Network for Calculating Functional Connectivity from Functional MRI', Medical Image Computing and Computer Assisted Intervention Society (MICCAI): Connectomics in NeuroImaging (CNI). doi: 10.1007/978-3-319-67159-8_9