foreword

After many years, it went online again and re-operated csdn. Since I was in graduate school, I have been either cooking or showing up. It is really decadent. It is not easy to start blogging again. Set a flag here, and strive to write something every week in the future. First, to practice writing, and second, to record the learning process.
My research direction is related to functional magnetic resonance imaging fMRI. I have never been in this field before. Start for nothing. After several months of study, although I am still confused, I can generally understand what it can do. In order to facilitate future study, the following simple knowledge points are summarized and organized

1. fMRI basics
2. fMRI data preprocessing
3. fMRI data analysis
4. Brain network connection analysis

1. Basics of fMRI

1. Brain diseases : Alzheimer's disease AD, mild cognitive impairment MCI, autism spectrum disorder ASD, depression MDD, ADHD, migraine, Parkinson's syndrome, schizophrenia, epilepsy, etc.

2. Brain imaging technology : computed tomography CT, magnetic brain imaging MEG, electroencephalogram EEG, positron emission tomography PET, functional magnetic resonance imaging fMRI, single electron tomography SPECT, magnetic resonance imaging MRI

3. Distribution and function of brain regions

Frontal lobe : including precentral gyrus, dorsolateral superior frontal gyrus, and middle frontal gyrus, affecting intelligence, cognitive ability, emotional management and behavior management
Hippocampus : Located between the thalamus and the temporal lobe, it affects human beings for long-term learning, such as processing and analysis of sound and light-related events, and emotional regulation processing
Right fusiform gyrus : Located in the temporal lobe, affects face recognition
Parietal lobe : including the postcentral gyrus, superior parietal gyrus, and precuneus, affecting the ability of humans to integrate information, such as language information
Top Back : Affects the ability to retrieve certain previous scenarios
Precuneus : Affects certain higher cognitive functions, such as the ability to perceive language, the ability to process self-related information, and the ability to remember scene information
Medial Temporal : Located in the temporal lobe, affects perception and inner and mental abilities
Temporal pole: middle temporal gyrus : a part of the temporal pole, which affects the perception and understanding of humorous language, events, and the ability to process emotions and society
Cerebellum : Affects reflexes and the ability to maintain body balance

4. Functional Magnetic Resonance Imaging fMRI

(1) Introduction : It is an imaging technique that obtains the functional structure of living organisms by measuring the changes of neuron activity in the living organisms and the oxygen level in the surrounding blood, combined with magnetic resonance imaging technology. Because it is a brain functional imaging technology based on blood oxygen level dependence, it is also often called a brain functional magnetic resonance imaging technology based on BOLD signal BOLD-fMRI

(2) Principle

Magnetic resonance imaging : Most nuclei have spin characteristics, and their spin axes are arranged randomly. If an external magnetic field is applied, the spin axes will change from a disordered state to an ordered state, and finally reach equilibrium. At this time, if an RF pulse of a certain frequency is added to the outside world, it will cause the resonance effect between the atomic nucleus and the magnetic field, resulting in a low-energy level transition, that is, the nuclear magnetic resonance phenomenon. When the RF pulse is withdrawn, the spin axis of the nucleus will return to its original state and release the energy. This process is called relaxation, and the relaxation is divided into vertical relaxation (T1) and horizontal relaxation (T2). There are two types of magnetic resonance imaging, one is structural image, which is the collection of magnetic resonance signals of specific nuclei in each part of the brain; the other is functional image, which is the collection of BOLD signals
BOLD functional imaging : When the brain is stimulated by the outside world, neurons need more oxygen to support their activities. This process will cause the BOLD signal of the relevant brain area to be significantly higher than that of other areas, and T2 will increase, making the BOLD signal of the corresponding brain area Significantly increased compared to unstimulated. The BOLD signal can be regarded as the result of the convolution of neural activity and the hemodynamic response function HRF, but it has a delay of several seconds compared with the neural activity, and it cannot directly reflect the neural activity of the brain, but indirectly reflects the neural activity by measuring the local blood oxygen level

(3) Features

Advantages : non-invasive, non-invasive, non-radioactive, no side effects, high safety, reproducible acquisition, simple acquisition process, high spatial resolution, high temporal resolution, precise positioning, clear imaging, and display of brain function activation Region location, size, extent and anatomical location
Disadvantages : The operating environment is harsh, and on-site observation is difficult. The delay and hysteresis of the BLOD mechanism will affect the accuracy of the experimental results. Due to the inhomogeneity of the magnetic field, small brain tissues may not be displayed normally.

(4) Applications : mental and physical diseases, brain cognition and other fields

2. fMRI data preprocessing

1. fMRI data : 4-dimensional data, including 3-dimensional spatial information and 1-dimensional temporal information. At present, the callback planar series imaging EPI method is mostly used to acquire fMRI functional images. The basic parameters of acquisition include repetition time TR, echo time TE, number of slices, spatial resolution, acquisition time, etc. Due to the fast acquisition speed of EPI, the time resolution is low, and the signal-to-noise ratio is also low. Although the spatial resolution of fmri data is high, as it increases, the voxel gradually becomes smaller, which will reduce the signal-to-noise ratio, so the voxel is generally set to 3×3×3mm

Repetition time TR : the time interval between two adjacent executions of the pulse sequence, representing the time required to scan a complete brain
Echo time TE : Indicates the time required from the first radio frequency pulse to the generation of the echo signal
slice number : the total number of slices obtained by scanning the brain
Volume : A reconstructed brain is a volume
Time Point : Usually a volume is a time point
Voxel : A voxel is a unit. The brain is divided into many small squares, which are many voxels. Usually, the resolution of the functional image is 3mm×3mm×3mm
Region of interest (ROI) : the brain area of interest

2. Public data sets

ADNI : Alzheimer's Disease Neuroimaging Initiative ( https://adni.loni.usc.edu/ )
ABIDE : Autism Brain Image Dataset ( http://fcon_1000.projects.nitrc.org/indi/abide/ )
Openneuro : Contains datasets of normal subjects, schizophrenia, bipolar, affective disorders, ADHD, etc. ( https://openneuro.org )

3. Software package

SPM : A function based on matlab, mainly used to analyze brain imaging data sequences, suitable for data analysis of fMRI, PET, EEG and MEG, analysis processing includes two parts: data preprocessing and statistical analysis (https: //www.fil .ion.ucl.ac.uk/spm/software/download/ )
DPARSF : A toolkit for processing and analyzing fMRI data based on SPM, which provides all data preprocessing operations and some common data analysis and processing functions, and can realize multifunctional automatic processing and batch processing functions ( http://rfmri.org/ DPARSF )
MRIcro : Brain functional image analysis and viewing software. Used to view and analyze fMRI data, including adjusting different time points, viewing corresponding images, setting thresholds to divide high signal values and low signal values, and visual display ( http://www.mricro.com )
BrainNet Viewer : A brain network drawing tool based on graph theory, which can display the nodes and edges in the brain network in the form of a graph, and combine the structural images of the brain to visually display the distribution of nodes in the brain network in the brain, as well as the nodes connection between. It not only supports voxel-level brain surface volume rendering, but also supports various commonly used division templates as node position rendering ( http://www.nitrc.org/projects/bnv/ )
FSL : a full-featured fMRI, MRI and DTI data analysis tool ( http://www.fmrib.ox.ac.uk/fsl/ )
WFU_PickAtlas : A software package that integrates the positioning of human brain functional regions and the generation of human brain ROI templates ( http://fmri.wfubmc.edu/ )

4. fMRI data preprocessing : perform a series of corrections, denoising, and unnecessary distortion removal on the image data obtained in the experiment
insert image description here

Format Conversion : Convert DICOM files to NIFIT format
Elimination of time points : In order to eliminate the influence of the fluctuation of the subject's state as much as possible, the data of the first 5 or 10 time points of the subject are generally eliminated
Slice Timing : Since a volume is scanned layer by layer, there is a certain time difference between the scanned images, and it is impossible to obtain the whole brain voxel signal at the same time point. Therefore, it is necessary to correct the time of different layers to eliminate the difference caused by different acquisition times, that is, to correct each slice data of each 3D MRI data to the same time point
Head movement correction (Realignment) : During the signal acquisition process, it is unavoidable to shake the subject's head due to breathing, heartbeat, etc., which will cause the scanned voxel position to change, resulting in inaccurate detection results. Therefore, through the rigid body transformation technology, the images collected in the time series are registered with the reference time points, so that the neighbors of the images do not move or control their movement range within a certain threshold
Image registration (Registrarion) : Due to the low resolution of functional images, it is difficult to analyze the specific information of brain regions. Therefore, registration of low-resolution functional images to high-resolution anatomical maps allows for comparability between different imaging
Normalization : Due to differences in the shape and structure of the brains of different subjects, the acquired image data is not in the same space, which is meaningless for subsequent experiments and analysis. Therefore, it is necessary to map the collected brain images of different subjects to the same MNI space to eliminate the differences caused by different brain shapes, so that the subjects are comparable
Spatial smoothing (Smooth) : Due to many irresistible external factors in the data collection process, there is a certain deviation between the final collected data and the real data, which affects the final analysis results. Therefore, it is necessary to smooth the data, which can eliminate the subtle errors caused by registration, standardization, etc., help to meet the Gaussian random field requirements in the statistical inference process, and improve the signal-to-noise ratio of the data and the validity of the parameter statistical test.

3. fMRI data analysis

1. Experimental design

Block design mode : the design scheme is simple, the operation is convenient, the obtained BOLD signal is stronger, and the data processing is easier, but the time curve of the BOLD signal of a single stimulus in the block cannot be obtained
Event-related design mode : It can detect the average BOLD curve caused by a single stimulus, and at the same time obtain signal data with a high signal-to-noise ratio, but the experimental design is more complicated and the data processing is more inconvenient
Resting-state design mode : It does not require the subjects to do any cognitive tasks or to do any additional thinking. The experimental design is not complicated. Appropriate ROI extracts meaningful time series, reflects the changes in signal intensity of brain voxels during the acquisition process, and finally obtains a time series curve that presents specific rules and organizational fluctuations over time
Tasking-state design mode : During the experiment, the subjects will be subjected to external stimuli set by the experiment or caused by the execution of the task, and the images collected are the changes in the blood oxygen level caused by the neural activity of the human brain

2. Research classification

Functional separation : the brain regions stimulated by different specific human activities do not overlap, different regions of the brain have different functions, are responsible for different activities, and do not need to cooperate with other brain regions
Functional integration : human behavior and cognition are not only supported by a single function, but are completed by the cooperation of the whole brain

3. Evaluation indicators : ROC curve, skewness analysis, correlation analysis of variance

4. Brain Network Connection Analysis

4.1 Functional connection

1. Introduction : Functional connectivity studies the information processing and transmission between functionally coordinated brain networks, which can reflect the way and content of communication between brain regions at a specific time, and provides a basis for studying information transmission and division of labor between different brain regions. method. It measures the correlation between brain regions and has no directionality. It generally reflects the strength of the relationship between brain regions by calculating the signals of different regions of the brain. Functional connectivity can only show functional connectivity and does not indicate a causal relationship between two brain regions, or even any direct connection between two brain regions. Various modalities of data and analysis methods can be used to assess functional connectivity, such as task-state fMRI data to analyze task-specific stimuli for functional connectivity, and time-series correlation of rest-state fMRI data to assess functional connectivity in the brain

2. Static functional connection SFC : It can measure the time correlation of spontaneous BOLD signals in different brain regions in the brain, and assume that regions with related activities will form a functional network, and believe that the functional connection between brain regions does not change over time change by change. Common research methods include model-driven, data-driven and functional network connectivity analysis

(1) Model-driven method : It is suitable for data with very definite experimental design parameters, and can obtain brain activation regions related to experimental tasks more conveniently and accurately, but often requires a large amount of prior knowledge and assumptions as a basis. Commonly used methods include correlation analysis method CCA, consistency analysis method CA and statistical parametric map analysis method SPM, etc. Among them, the correlation analysis method and the consistent analysis method need the prior information of the precise positioning of the ROI, while the statistical parametric map analysis needs information about the stimulation pattern. prior information, which would limit access to functional connectivity networks from a global perspective

Correlation analysis method CCA : The degree of correlation between two brain regions is analyzed through the correlation coefficient. When the correlation coefficient is greater than a certain threshold, it is considered that there is functional connectivity between the two brain regions. The commonly used measurement methods are Pearson correlation, log likelihood similarity and mutual information. Taking Pearson correlation as an example, its analysis process is (1) determine the location of the brain region or voxel contained in the ROI (2) obtain the time series of the ROI (3) use the Pearson correlation coefficient to calculate the relationship between the ROI and the brain region Corresponding coefficient (4) for further analysis of the obtained correlation coefficient. The disadvantage of this method is that it needs to manually select the seed point/ROI according to the prior knowledge of neurocognition, and the selection of ROI is accidental and random.
Consistency analysis method CA : The cross-correlation between time processes in the time domain is transformed into the consistency analysis between corresponding power spectra in the frequency domain. To a certain extent, it can eliminate the deficiency that the cross-correlation analysis method is susceptible to physiological noise, but this method is still subject to the influence of random selection of seed points or regions of interest.
Statistical parameter map analysis method SPM : take each voxel of brain functional imaging data as the basic unit, and perform statistical analysis voxel by voxel to obtain a brain function activation map at a certain level of significance, and calculate some Temporal or task-dependent curves of physiological signals in (or all) activated regions

(2) Data-driven method : multivariate analysis method, without relying on any experimental model analysis method, commonly used methods include principal component analysis, independent component analysis and cluster analysis

Principal Component Analysis PCA : Extract the main components of the data according to the maximum variance, and then use the main components for feature extraction and classification prediction. Usually, the basis vector whose dimension is smaller than the original data can be used, so that the original high-dimensional data can be projected into the low-dimensional space to achieve the purpose of dimensionality reduction, thereby removing redundant values in the data, eliminating data noise, reducing the feature dimension of the data, and in the remaining The characteristic information of the original data is preserved in the dimension
Independent Component Analysis ICA : An effective component analysis technique developed from signal blind source analysis. Using a mixture matrix and an independent composition matrix to represent the experimental data, the effective composition can be considered as a network. The resulting components are often transformed by z_score, and the higher the z score, the stronger the functional connection between corresponding voxels. This method does not predefine the ROI, but separates multiple components through the original data, and then calculates the functional connections between the components
Cluster analysis method CA : According to a certain similarity rule (Euclidean distance, Pearson distance), similar voxels are gathered together to form multiple clusters representing different brain functional connection networks, usually clustering Each class obtained is called a state of the brain. Commonly used methods are fuzzy clustering FCA, hierarchical clustering HC and affine clustering APC

(3) Functional network connection analysis method FNC : comprehensive data-driven and model-driven. Similar to ROI-based methods, but this method can also reflect differences in brain regions and connections between brain regions over time. The calculation steps are (1) extract the independent components of the subjects by independent component analysis (2) decompose the time series of the components by using various methods such as space-time regression (3) calculate the average of each network component time series (4) calculate Functional connectivity between networks

3. Dynamic functional connectivity (DFC) : functional connectivity has special temporal dynamics, which can capture information that cannot be detected by static functional connectivity, better reflect the temporal characteristics of functional connectivity, and reveal the information exchange between brain regions. Commonly used methods are sliding window method and windowless method

(1) Sliding window method

Steps : (1) Determine the length of each sliding window (30s-60s is appropriate) and step size (2) Calculate the correlation coefficient of the time series between voxels or brain regions contained in each window (3) Correlation The coefficients are z-scored and used as the functional connectivity matrix (symmetric matrix) of the window (4) Take the upper (lower) triangle of each window functional connectivity matrix and convert it into a column vector, and stitch these column vectors into a functional connectivity matrix by row (5) Perform K-means clustering on all windows to obtain a stable pattern of dynamic functional network connectivity
Limitations : Different window lengths will lead to different research results. If the window is too short, the data contained in the window is very limited, and the calculated connection is not stable enough to accurately represent the dynamic functional connection within a period of time; if the window is too long, it is close to the static functional connection and cannot reflect the time characteristics of the dynamic functional connection

(2) Windowless method : It can effectively solve the problem of lack of uniform standard for window length, but still needs further exploration. The commonly used method is time-frequency analysis, that is, to adapt the observation window to the frequency of the time course

4.2 Effective connections

1. Introduction : To measure whether the activity of one brain area will change due to the change of the activity of another brain area. It is characterized by dynamics and causality. In the resting state, it can reflect the information flow inside the brain without external stimuli. , while considering the interaction between nodes. Effective connections are directional. It not only considers the time factor, but also considers the direction of information transmission between brain regions. Common methods include Granger causality model GCM and dynamic causality equation model DCM. Using effective connections can extract more features of brain networks, thereby improving the depth of brain network research

2. Method

Granger causality analysis (GCA) : causal analysis can be performed at the whole brain level without assuming the existence of anatomical connections between the study regions in advance to form a GCM model, and it is more convenient to select ROI. The disadvantage is that it cannot identify the weak-coupling to medium-coupling relationship between variables very well, and is easily affected by noise
Cross Convergent Mapping CCM : Based on the theory of state space reconstruction and Takens embedding theory, using the diffeomorphism characteristics between the reconstruction spaces of variables, the causal relationship between variables can be judged by the mutual prediction of the reconstruction spaces. More suitable for inseparable dynamical systems similar to the human brain

4.3 Complex Networks

1. Graph theory : an effective tool for analyzing complex networks, provides a powerful data structure framework for quantifying the structure of complex systems, and provides accurate methods and indicators for analyzing the topological characteristics of complex systems

2. Topological attribute analysis

(1) Clustering coefficient : An index used to describe and quantify the degree of node aggregation. For a complex network, the node aggregation coefficient can indicate the degree of connection aggregation between a node and other surrounding nodes. The degree of network aggregation refers to the degree of mutual clustering between nodes in the entire network.

(2) Shortest path length : It describes how a node in the network can reach another node with the least cost or maximum efficiency

(3) Degree : One of the most basic attributes of a graph. The greater the degree of a node, the more nodes connected to the node. A type of node selected by a certain standard, such as the degree value greater than the average + standard deviation of all node degrees, is called a central node.

(4) Centrality : Describe the importance of nodes in the network. The greater the centrality value of a node, the more important the node is. The centrality can be defined by the degree of a node, that is to say, the greater the degree of a node, the greater its centrality. Centrality can also be defined from the perspective of information flow, that is, betweenness centrality

(5) Isomorphic type : used to examine the tendency of interconnection between nodes with relatively close degrees

(6) Modular structure : refers to the clusters of nodes that are sparsely distributed among each other but closely connected internally in complex networks. It is mainly divided into two modules, which are tightly connected internally and sparsely connected to each other. The nodes with more connections inside the module are hubs in the network, called hub nodes, and the nodes with more connections between modules are connected hubs

(7) Local and global efficiency

3. Network model

(1) Small-world network : a network between regular networks (higher clustering coefficients, longer shortest paths) and random networks (lower clustering coefficients, shorter shortest paths), that is, shortest paths at the same time and networks with higher clustering coefficients

(2) Scale-free network : a network showing high robustness and vulnerability

fMRI basic theoretical knowledge learning