论文全名:Functional Connectivity Network Analysis with Discriminative Hub Detection for Brain Disease Identification
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
2.3.1. Network-based brain disease diagnosis
2.3.2. Hub detection from networks
2.4.2. Connectivity network analysis with discriminative hub detection (CNHD)
2.4.4. Computational complexity
2.5.1. Experimental results and analysis
1. 省流版
1.1. 心得
(1)没推公式,感觉就是把特征选择和分类都集合到损失里去了?
1.2. 论文总结图
2. 论文逐段精读
2.1. Abstract
①Existing problems: a) extracted feature from brain connectivity might not suitable for the subsequent classifier; b) cannot exactly capture the hubs
②They prorposed a Connectivity Network analysis method with discriminative Hub Detection (CNHD) for brain disease diagnosis
2.2. Introduction
①Further explain the two existing problems
②They unify feature extraction and classifier together
③They developed an alternating direction method of multipliers (ADMM) to solve the proposed nonconvex optimization problem
④Datasets: 3 SCZ datasets
2.3. Related work
2.3.1. Network-based brain disease diagnosis
①Topolology-based methods: using graph kernal or clustering coefficient to capture features and send them to classifier
②Subgraph based methods: using different subgraph to represent each graph
immunodeficiency n. [免疫] 免疫缺陷
2.3.2. Hub detection from networks
①Listing several methods which aim to find the hubs
2.4. Methodology
2.4.1. Notations
①Introducing some common notations in this paper
2.4.2. Connectivity network analysis with discriminative hub detection (CNHD)
①For subjects with
ROIs/nodes each, the network set is defined as:
where denotes the label of each participant.
, -1 denotes the healthy samples and 1 denotes the diseased one
②The prediction goal:
where is the empirical loss term,
is the regularization term. The weight matrix
is a symmetric matrix with all elements in diagonal is 0
③The logic loss function:
denotes the bias term
④They decomposite to sparse symmetric matrix
and
whose "columns are either entirely zeros or entirely non-zeros":
(哪里整列0或者整列非零了?). "sparse entries of
represent edges between non-hub nodes, and the non-zero columns of
are regarded as hub nodes."(我没能非常get)
⑤CNHD module:
where denotes non-negative parameters
2.4.3. Optimization algorithm
①这,好难打字,我截图了(这个是改变上面⑤式子的约束):
②They change the original loss function to:
where denotes the Lagrange multiplier,
is the penalty parameter.
③Iteration steps:
2.4.4. Computational complexity
①Time complexity of their algorithm:
②Overall algorithm:
2.5. Experiments
(1)Datasets
①Affiliated Nanjing Brain Hospital of Nanjing Medical University (NBH): 24 SZ and 21 NC
②Center for Biomedical Research Excellence (COBRE): 67 SZ and 53 NC
③National Taiwan University Hospital (Taiwan): 69 SZ and 62 NC
④Atlas: AAL-90
⑤FC construction: Pearson
(2)Competing methods
①Introducing compared methods
(3)Evaluation metric
①Compared metrics: ACC, SEN, SPE, balanced accuracy (BAC), positive predictive value (PPV), negative predictive value (NPV), AUC
(4)Experimental settings
①Validation: 5-fold cross-validation with 4 for training and 1 for testing
②Grid search for hyper-parameters:
2.5.1. Experimental results and analysis
(1)Results of brain disease classification
①Comparison table:
(2)Analysis of detected network hubs
①Hubs in one fold:
②Visualized hubs:
(3)Convergence evaluation
①Iteration figure:
2.6. Conclusion
Their model unify the feature selection and classifier, achieving a excellent performance,
3. 知识补充
3.1. L2,1范数
(1)定义:L2,1范数定义为矩阵每一行的L2范数之和
(2)举例:
4. Reference
Wang, M. et al. (2019) 'Functional Connectivity Network Analysis with Discriminative Hub Detection for Brain Disease Identification', AAAI. doi: 10.1609/aaai.v33i01.33011198