论文全名:An overview of artificial intelligence techniques for diagnosis of Schizophrenia based on magnetic resonance imaging modalities: Methods, challenges, and future works
论文网址:基于磁共振成像模式的精神分裂症诊断人工智能技术概述:方法、挑战和未来工作 - 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
2.4. AI methods for SZ detection
2.5. CADS for schizophrenia diagnosis based on artificial intelligence methods
2.5.1. Available MRI modalities datasets
2.5.2. Preprocessing methods for sMRI and fMRI modalities
2.5.3. Artificial intelligence methods
2.6.1. Comparison of ML and DL papers accepted for SZ detection in each year
2.6.2. Number of MRI datasets used in SZ research every year
2.6.3. Number of MRI neuroimaging modalities used in SZ research of every year
2.6.4. Preprocessing toolboxes for diagnosis of SZ
2.6.5. Number of classification algorithms in ML and DL methods for SZ diagnosis
2.7.1. Challenges in MRI datasets
2.7.2. Challenges in multimodality datasets for SZ detection
2.7.3. Challenges in SZ detection
2.7.4. Challenges in ML techniques
2.7.5. Challenges in DL techniques
2.8.1. Future works in datasets
2.8.2. Future works in multimodality neuroimaging datasets for SZ detection
2.8.3. Future works in ML methods
2.8.4. Future works in DL methods
2.8.5. Future works in the SZ treatment methods
2.8.6. Future works in rehabilitation
2.9. Conclusion and future works
3.1. Preferred reporting items for systematic reviews and meta-analysis (PRISMA)
3.3. Out of distribution generalization
1. 省流版
1.1. 心得
(1)我第一次看到如此的七层嵌套,导致我突然做笔记都不知如何下手了
(2)特征选择那介绍的太详细了,或者可以不分点分那么多。
1.2. 论文框架图
2.论文逐段精读
2.1. Abstract
①Phenotypic symptoms of Schizophrenia (SZ): abnormal behavior, perception of emotions, social relationships and reality perception
②Physiological symptoms of SZ: changes of temporal and anterior lobes of hippocampus regions
hallucination n.幻觉;幻视,幻听(尤指生病或毒品所致);幻象;幻视(或幻听)到的东西
2.2. Introduction
①Possible etiology: genes and several environmental factors
②Distribution of SZ all over the world
③Categories of SZ:
④Diagnosis methods of SZ: electroencephalography (EEG), magnetoencephalography (MEG), functional near-infrared spectroscopy (fNIRS), and functional MRI (fMRI)
⑤Challenge of EEG: precisely localizate active brain regions
2.3. Search strategy
①Database: IEEE Xplore, ScienceDirect, SpringerLink, and Wiley
②Keywords: “Schizophrenia”, “sMRI”, “fMRI”, “Machine Learning”, “Feature Extraction”, “Artificial Intelligence”, and “Deep Learning”
③Progress of filtering:
④Criteria of filtering:
2.4. AI methods for SZ detection
①Reviews comparison:
2.5. CADS for schizophrenia diagnosis based on artificial intelligence methods
①The processing flow of ML and DL for SZ diagnosis:
2.5.1. Available MRI modalities datasets
①Dataset: Schizconnect, northwestern university SZ data and software tool (NUSDAST), center for biomedical research excellence (COBRE), function biomedical informatics research network (FBIRN), MCIC, University of California, Los Angles UCLA, and IEEE international workshop on machine learning for signal processing 2014 (MLSP 2014)
(1)Schizconnect
①Sample: 1392 with 632 NC, 215 broad SZ, 384 strict SZ, 41 schizoaffective disorder, 10 bipolar disorder,44 sibling of SZ strict and 66 sibling of no known disorder(这也太杂了)
(2)NUSDAST
①Sample: 450
②Categories of image: sMRI, landmarks maps, FreeSurfer measurement, and segmentation
③Cognitive data: scores for crystallized intelligence, working memory, episodic memory, and executive performance
④Clinical data: demographics, sibling relationships, the scale for assessment of positive symptoms (SAPS), and negative symptoms (SANS) psychopathology
⑤Genetic data: 20 single nucleotide polymorphisms (SNPs)
(3)COBRE
①Modalities: rs-fMRI and sMRI
②Sample: 72 SZ and 75 NC
(4)FBIRN
①Modality in stage II: T1-weighted and T2-weighted sMRI(这篇论文又合并在一起了...)
②Sample in stage II: 87 DSM-5 SZ or schizoaffective disorder and 85 NC
③Modalities in stage III: sMRI, fMRI,DTI, and behavioral data
④Sample in stage III: 176 SZ and 186 NC
(5)MCIC
①Modalities: sMRI, fMRI, and DTI
②Sample: 162 SZ and 169 NC
③Assessments: clinical and cognitive assessments, genetic testing, etc
(6)UCLA
①Modalities: T-fMRI, rs-fMRI, sMRI, diffusion-weighted imaging (DWI), and phenotype information
②Sample: 50 SZ, 49 bipolar disorder, 43ADHD, and 130 NC
(7)MLSP 2014 Schizophrenia classification challenge
①Modalities: sMRI and fMRI
②Sample: 69 SZ and 75 NC
2.5.2. Preprocessing methods for sMRI and fMRI modalities
①Preprocessing toolkits: FMRIB Software Library (FSL), brain extraction tools (BET), FreeSurfer and statistical parametric mapping (SPM)
(1)Standard (low-level) sMRI preprocessing steps
①Denoising: iscrete-time wavelet transforms (DWT)
②Inhomogeneity Correction: mitigating impact of coil damage on MRI equipment
③Skull-stripping
④Registration: register to MNI standard space
⑤Intensity standardization: histogram matching techniques and methods
⑥De-oblique: record all cerebral tissue by oblique scan
⑦Re-orientation
⑧Segmentation
⑨Flow chart of sMRI preprocessing:
(2)Standard (low-level) fMRI preprocessing steps
①Removing: removing first N volumns adaptively
②Slice timing: correct time consistency of different slices
③Head movement correction: balancing the data by a reference image in the sept of Motion Correction and Volume Scrubbing
④Normalization: pattern standardization and spatial normalization, such as MNI
⑤Spatial smoothing: weighted mean value of BOLD signals of the neighboring voxels
⑥Temporal filtering: 0.08–0.009 Hz band-pass filter
⑦Flow chart of fMRI preprocessing:
2.5.3. Artificial intelligence methods
(1)Conventional machine learning methods
①Details of SZ diagnosis using different ML models:
②Feature extraction techniques in ML:
Statistical | mean, variance, standard deviation, moments etc. |
Texture | Gray level co-occurrence matrices (GLCM)-based methods and Gabor filters |
Non-Linear | |
Graph | construct or extract graph representation |
Connectivity Matrix | Functional connectivity matrix (FCM) and structural connectivity matrix (SCM) |
③Feature reduction/selection techniques in ML:
Feature reduction techniques | principal component analysis (PCA) |
Feature selection methods | a) supervised: Relief, Fisher, Chi-Squared, and correlation, b) unsupervised: variance, mean absolute value of the differences, scatter ratio, Laplacein score and clustering, c) optimization: Genetic algorithms (GA), ant colony optimization (ACO), binary particle swarm optimization (BPSO), and non-dominated sorting genetic algorithm II (NSGA-II) |
agglomerative adj.凝聚(结)的;附聚的;烧结的;胶凝的
(2)Deep learning methods
①Details of SZ diagnosis using different DL models:
②Introducing CNNs (1/2 D CNN, Inception, Generative Adversarial Networks (GANs), CapsNet, 3D CNN), RNNs (Long short term memory (LSTM) and gated recurrent units (GRU)), Autoencoders (AEs) (stacked AE, denoising AE, sparse AE), Deep belief networks (DBNs), CNN-AE, CNN-RNN
(3)Classification methods
①Classification approaches: the support vector machines (SVM) (DL/ML), random forest (RF) (DL/ML), and Softmax (DL)
2.6. Discussion
2.6.1. Comparison of ML and DL papers accepted for SZ detection in each year
①Number of ML/DL method based papers on SZ diagnosis:
2.6.2. Number of MRI datasets used in SZ research every year
①Number of cited dataset on 2016-2021:
2.6.3. Number of MRI neuroimaging modalities used in SZ research of every year
①Number of paper using different modalities on 2016-2021:
2.6.4. Preprocessing toolboxes for diagnosis of SZ
①Number of used toolkits:
2.6.5. Number of classification algorithms in ML and DL methods for SZ diagnosis
①Classifiers used in ML and DL during 2016-2021:
2.7. Challenges
2.7.1. Challenges in MRI datasets
①Limited number of data
②Low resolution ratio on spatial sMRI, DTI, and temporal-spatial fMRI
③Limited categories of SZ
④Limited longitudinal data
2.7.2. Challenges in multimodality datasets for SZ detection
①The complexity of multimodality
2.7.3. Challenges in SZ detection
①Similar symptom of different diseases
2.7.4. Challenges in ML techniques
①Hard to select features accurately
②Large amount data is hard for ML method
2.7.5. Challenges in DL techniques
①Limited dataset
②Hardware resources
2.7.6. Challenges in hardware
......(其实已经没有必要再单拎出来说一下了,和上面重复了)
2.8. Future works
2.8.1. Future works in datasets
①More patients
②Various ages
2.8.2. Future works in multimodality neuroimaging datasets for SZ detection
①Multi-modal dataset required(所以这个前面也提到了啊,有点重复了)
2.8.3. Future works in ML methods
①Applying fuzzy clustering(啊??这么细吗?这么绝对吗)
②后面已经不想做笔记了,为什么要去建议用什么方法?方法不都是实时更新的吗?
2.8.4. Future works in DL methods
(1)Attention models
①......怎么,注意力一统天下了?
(2)Out of distribution generalization
①Constructing a unified robust model for different dataset with different distribution
(3)Robust representation learning
①Augmenting data representations by a) finding different concepts in vectors, b) learning to disentangle the representation, c) finding a causal representation
(4)Generative adversarial nets (GANs)
①CycleGAN, deep convolutional GAN (DCGAN), and Info-GAN are used for SZ diagnosis
(5)Deep compact-size CNNs
①Deep compact-size CNNs such as Facebook-Berkeley-Nets (FBNet), PP-LCNet, MnasNet, MobileNet, and TinyNet, which cost less for training, are used for diagnosis as well
2.8.5. Future works in the SZ treatment methods
①Clinic diagnosis methods of SZ: transcranial magnetic stimulation (TMS) and transcranial direct current stimulation (tDCS)
2.8.6. Future works in rehabilitation
①CADs
2.9. Conclusion and future works
讲了一下哪些段落或者哪些表格做了什么。也没必要写这么长吧。
2.10. Appendix A
①Evaluation metrics and corresponding formula:
2.11. Appendix B
①Abbreviations from A to Z
3. 知识补充
3.1. Preferred reporting items for systematic reviews and meta-analysis (PRISMA)
(1)网址:棱镜 (prisma-statement.org)
(2)介绍:
PRISMA是一套用于系统评价和荟萃分析(meta-analysis)的报告指导方针。它提供了一套项目清单和流程图,以帮助作者更透明和完整地报告他们的研究结果。PRISMA广泛应用于医疗领域,被认为是报告系统评价和元分析的黄金标准。
PRISMA的主要目的是确保系统评价和荟萃分析以透明和全面的方式进行和报告,从而提高研究结果的可信度和有效性。它帮助作者明确列出应该包括在项目中的关键元素,并说明研究选择的过程,以确保过程和发现是透明的,并且对其他人来说是可理解的。此外,PRISMA还检查表和流程图帮助作者确保所有相关信息都包含在他们的系统评价或荟萃分析中。
值得注意的是,PRISMA并不是一种进行系统评价或元分析的方法,而是一套用于报告这些分析的指导方针。在实际操作中,PRISMA表是进行元分析时的必要材料,用于汇总、比较某一课题下的多项研究的结果,并通过如森林图等方式直观地展示研究结果。
(3)展示:
3.2. Fuzzy clustering
参考学习1:三分钟搞懂模糊聚类(Fuzzy Clustering Mean ,FCM) - 知乎 (zhihu.com)
参考学习2:机器学习之模糊聚类——FCM,PCM模型_pcm模糊聚类算法实现-CSDN博客
3.3. Out of distribution generalization
(1)介绍:
分布外泛化(Out-of-Distribution Generalization,简称OOD泛化)是机器学习中的一个重要概念,旨在解决训练数据与测试数据分布不一致的问题。简单来说,它要求模型在有限的观测数据基础上,学习出背后稳健的规律与关联性,并能够将这些规律和关联性应用到未知的数据分布中。
在机器学习的传统框架中,大多数模型都是基于独立同分布(Independent and Identically Distributed,I.I.D.)的假设进行训练和测试的。然而,在实际应用中,训练和测试数据的分布往往存在差异,这就导致了模型在测试数据上的性能下降。OOD泛化正是为了解决这个问题而提出的。
为了实现OOD泛化,研究者们提出了各种方法和技术。其中,一种常见的方法是通过给每个样本加权,来消除虚假关联,使得模型更加关注与分类结果有因果关系的特征。这样,即使训练数据和测试数据的分布存在差异,模型也能准确地识别出与标签有因果关系的特征,从而提高泛化能力。
目前,OOD泛化已成为机器学习领域的一个研究热点,为图学习等子领域引入了新的研究方向。随着研究的深入,OOD泛化有望在解决现实世界的复杂问题中发挥更大的作用。
(2)相关论文:[2108.13624] Towards Out-Of-Distribution Generalization: A Survey (arxiv.org)
4. Reference List
Sadeghi, D. et al. (2022) 'An overview of artificial intelligence techniques for diagnosis of Schizophrenia based on magnetic resonance imaging modalities: Methods, challenges, and future works', Computers in Biology and Medicine, 146. doi: Redirecting