Nature Neuroscience: challenges and future directions of functional brain tissue characterization

Abstract ： A key principle of brain organization is to integrate the functions of brain regions into interconnected networks . Functions acquired at rest MRI Scanning through coherent wave patterns in spontaneous activity , The so-called functional connection , Provides insight into functional integration . These patterns have been thoroughly studied , And related to cognition and disease . However , This field is subdivided . Different analysis methods will divide the brain into different categories , It limits the replication and clinical transformation of research results . The main source of this division is the method of simplifying complex brain data into low Witt collection for analysis and interpretation , This is what we call brain representation . In this paper , We provide an overview of different brain representations , List the challenges that lead to the segmentation of this field and continue to form convergence barriers , And put forward specific guidelines for unifying this field . 1. sketch Resting state MRI The research field of is classified , About the pretreatment process 、 Brain zoning method 、 The post-processing analysis method and endpoint are controversial . The main source of this problem is the challenge of brain representation . Magnetic resonance imaging produces a large number of high-dimensional data , A major analytical task is to extract interpretable content from the huge complexity of measured brain activity . Here we use “ Brain representation ” To describe the dimensionality reduction process . Brain representation is a collection MRI Multidimensional description of data , Including the spatial definition of brain units （ Partition ） And extracting the overall measure of interpretable features at the brain unit level （ If pairing is related ）. How to characterize brain data fundamentally establishes the description of brain function and organization . The representation of the brain is often considered as a mapping problem , It aims to eliminate the boundary of different areas of neuroanatomy of function and nerve tissue . However , Brain representation includes the form of representation and how data is transformed into these representations . This article aims to provide a framework for consistency and repeatability in this field rfMRI An introduction to characterizing challenges . 2. Introduction to brain representation Brain representation can be obtained by BOLD The data is reduced to a set of features for analysis . Many brain representations recognize ：1） A group of low dimensional brain units （ Space division ）2） A set of measures applied at the brain unit level （ Pairing correlation ）. These features are used for later statistical or predictive analysis . use “ Brain unit ” To denote a neural entity defined in any space , It can be regarded as a basic functional processing unit .“ Measure combination ” As a method of calculating characteristics , Relative to brain unit definition . Combinatorial measures are used to answer research questions , So it is relative “ Specific domain ” Of . A small number of brain representations do not use brain units and combinatorial measures , Instead, use estimated features , Complex spatiotemporal patterns that can represent activities . 2.1 Define a brain unit rfMRI Spatial resolution can easily reach 2x2x2mm³, This will get about in the whole brain 100000 Voxel .rfMRI in , These voxels （ Or the summit ） Is the smallest measurable brain unit . However, it does not represent a specific level of neuroanatomy . Therefore, voxels or vertex units will be combined into a smaller set of brain units to achieve meaningful low-level brain representation . Brain units may or may not be spatially adjacent . Adjacent brain units are consistent with specific functional cortical areas （ chart 1a）, Nonadjacent brain units can capture the complex network structure of hierarchical organization and large hemispherical symmetric brain （ chart 1b）. Brain units can be binary （ A voxel or vertex is assigned to a unit ） Of or weighted （ Voxels or vertices contribute to multiple cells according to their weights ）. There are many ways to define brain units . The obvious choice is based on histology 、 Pathological changes 、 Division of atlas defined by folds or other features . But these atlas come from a small number of people , And the anatomically defined boundary does not necessarily match the functional tissue . Many methods use functional data to define partitions , Include ICA,PCA, Nonnegative matrix decomposition , Probability function module or dictionary learning . This partitioning depends on spontaneity BOLD wave , Limits its applicability . Deconstruction with 、 Resting 、 The multimodal approach of task combination may provide more extensive zoning .

chart 1 Examples of brain representation 2.2 Define composite measures Functional connectivity combination measure ：rfMRI The most common type of information in research is functional connectivity , It is defined as the statistical similarity between signals from different brain regions , It is considered as an indication of functional integration . Graph based connectomics captures functional connectivity information by abstracting a single brain unit into nodes in a graph . Connectomics is the study of all possible pairs of node to node functional connections ( edge ), Can be summarized in matrix form （ chart 2a）. Functional connections can be temporal or spatial , This fuzziness leads to the challenge of brain representation interpretation . Functional connections can also be static or dynamic . Dynamic function connection can identify different dynamic states （ chart 2b）. Another variant is to infer cause and effect by estimating the directed connection from one brain unit to another （ chart 2c）. Univariate （ Node based ） The combinatorial measure of ： Although most rfMRI Combined measurements of brain representation assess functional connectivity in some form ( Integration ), But there are several alternative combination measures that describe various aspects of the data for each brain unit . Such as local signal amplitude （BOLD Strength ）, Brain unit size , Weighted brain unit spatial overlap . Although they are different in nature , These univariate measurements are usually not independent of functional connectivity . for example , The change of signal amplitude may directly affect the functional connectivity . Complex spatiotemporal brain representation ： Although all the methods discussed so far start with the definition of brain unit , But some brains say they avoided this step , And estimate complex spatiotemporal patterns from complete data . for example ,rfMRI The data can be represented as a ( Or more ) Connecting gradients , The gradient captures changes along the spatial position function connection of the continuous axis . This method can be used to identify overlapping patterns of tissues in predefined brain units , Or map the main global patterns of cortical organization from the primary sensorimotor cortex to the multimodal contact cortex （ chart 2d）. In addition, there are dynamic changes in time and space “ Propagating waves ” Brain representation of .

chart 2 Different combinations of measures in different brain representations are based on FC Version of 3. The challenge of brain representation at present rfMRI Brain representation of data ( That is, combine the description of brain units with the summary measurement ) The differences are natural and expected , As part of the initial exploration phase , It has been responded to in other disciplines . However , Now this field has matured to include biomarker discovery , To build a cumulative scientific framework , Efforts need to be made to integrate the presentation of validation . Due to the lack of gold standard to verify and compare brain representation , These efforts have become complicated . Non invasive techniques cannot completely capture the potential neural tissue of an individual . The comparison between different representations must depend on, for example, the accuracy of behavior prediction 、 Genetic heritability 、 Intra partition homogeneity 、 Variable interpretation 、 measurement - Retest reliability 、 Indirect indicators such as comparison with other modal data . In this section , We list some of the challenges of brain representation caused by the lack of basic truth knowledge , The aim is to raise awareness of issues that are rarely explicitly considered or taught in certain situations . 3.1 Heterogeneity and brain unit dimension A common assumption inherent in most brain representations is , A single brain unit is functionally homogeneous , Therefore, its related activities can be accurately reflected in a single summary time series . However , Functional heterogeneity within a brain unit can be measured by noise 、 Structured human facts 、 Variability between participants and true heterogeneity at the level of neural processing . In addition to heterogeneity , When the same part of the cortex encodes different types of information , Functional diversity also occurs . Examples of this diversity can be seen in the visual cortex , It encodes both retinal bitmap and stimulus orientation , Or in the parietal cortex , Different body position maps converge and overlap . There are inevitably some neuronal functional heterogeneity and / Or diversity , It is generally accepted that , Even hypothetical , But the meaning of brain representation is rarely considered or explained . One potential way to reduce the problem of functional homogeneity and diversity is to divide partitions into smaller brain units , To achieve a more granular representation of the brain . However , Overly fine segmentation , Multiple brain units are used to represent the same functional entity , May lead to complexity in modeling and interpretation . for example , If a functional area is improperly divided into multiple brain units , So when using partial correlation , It will lead to the wrong estimation of the functional connection , And have an adverse impact on the causal connection model . Determine the optimal number of brain units in brain representation , Balancing the trade-off between homogeneity and model complexity is a challenge . There is no consensus on the optimal dimension of brain representation , The most recent suggestions are from 6 There are macro scale systems to hundreds of subareas . This wide range is partly due to the hierarchical organization of the brain , According to the research problem, it can be meaningfully expressed at many different granularity levels . for example , The topological characteristics of brain functional organization can be studied in different dimensions , Different patterns of variability within and between participants may dominate on different scales . However , Due to improper handling of variability between participants , It will also lead to the increase of dimension estimation , This leads to misleading granularity of detail . It is worth noting that , Due to hemodynamic mediated BOLD The dependence of signal on the structure and latency of cerebral microvessels ,rfMRI The effective dimensions of data are biologically limited . The challenge of these heterogeneity shows , The best model of human brain and rfMRI There is a disconnect between the best models of measurement . There is sufficient evidence to show that , In humans and other species , There are neural groups with specific functions , These neural groups are organized into different cortical areas . According to this evidence , Binary segmentation into adjacent brain units may therefore be the best macro model of the brain . However , Although rapid progress has been made in recent years with the help of accelerated acquisition methods , The spatial-temporal resolution of fMRI has removed many orders of magnitude from the scale of neural population and action potential . Similarly , The physiology of hemodynamic response means the limitation of resolution , Progress independent of MRI image acquisition . therefore ,rfMRI The data obtained provides a rough measure of information in both space and time . therefore , The weighted segmentation that allows overlapping organization and fuzzy boundary can be rfMRI The measured data provides a better model . in fact , Previous comparisons have shown that , In predicting behavioral characteristics , Weighted brain representations may perform better than binary combinations . However , Care should be taken in interpreting this weighted segmentation . for example , Spatial overlap between brain units may be an important combinatorial measure to consider . 3.2 Deal with variability Brain representation is usually defined based on a large number of participants , To achieve consistency between individuals , Compare between groups , And overcome the limited signal-to-noise ratio in single participant data . However , The variability of brain functional organization measurement in different individuals may be due to the spatial imbalance between individuals or the real individual differences in brain structure / Or function . Although the surface based alignment method aims to solve this variable , But a recent study of widely scanned individuals suggests that , In group derived brain representations , The specific individual characteristics of the organization are distorted . Some recent methods aim to estimate personalized partition boundaries , Integrate group and participant estimates in the same Bayesian framework , To solve these problems of variability between participants , The natural movie viewing paradigm is adopted to control the variability in the data acquisition process or move towards connectivity based super alignment across participants . In addition to the differences between participants , as time goes on , The instability of brain representation in an individual ( for example ,sessions The difference of , Even individuals session Internal dynamics ) Is a further source of brain representation differences . Although some studies have reported stable class characteristics of brain representation , But other work shows state dependency changes based on task requirements , And the wake-up state 、 Physiologically related fluctuations . Within participants , With development 、 Longitudinal changes in aging or disease progression have not been identified . All in all , The potential sources of variability within these participants point to the importance of eliminating the effects of features and states in brain representation . Intra participant variability 、 The variability between participants and the complex interaction between dimensions constitute the main challenges to the definition and interpretation of brain representation . The application of brain representation is mainly to investigate the effects between participants ( for example , Patient control comparison ; Personal fingerprint recognition ; On behavior 、 Prediction or regression of cognition or diagnosis ). therefore , It is important to determine which brain representation is most sensitive to the effects between participants . for example , Estimating individualized brain unit boundaries , In order to eliminate the dislocation as the source of influence between participants , Will improve interpretability . Besides , It is very important to determine the best dimension through empirical comparison and summarize and measure the problems between specific participants . function MRI (fMRI) The non neural confusion in the data adds another unnecessary source of difference . Structural artifacts may be caused by the movement of the participant's head 、 Heart and respiratory cycles and these participant factors and magnetic fields ( Inside ) Homogeneity 、 The interaction between the excitation pulse and the image readout causes . At present, the pretreatment methods used to remove non neural mixtures are not perfect , Apply it to rfMRI Data can have unwanted side effects .rfMRI Summary measurements of brain representation are mostly based on measurements BOLD Signal similarity , Therefore, one or more sources of random variation are required ( for example , Spontaneous fluctuations in neural activity ). The uncertainty of these sources increases rfMRI The effect of confusion ( Usually more than a task fMRI More serious ). therefore , Developing and comparing improved data preprocessing strategies is an active research and discussion field . 3.3 Characterization fuzziness According to the research objectives of the given research , The fuzziness of these representations may be more or less crucial to the final conclusion . For functional brain tissue ( Such as the similarity between resting state and task organization ) A broader view of the brain should be relatively independent of the details of the chosen brain representation . Again , If the goal is to achieve accurate clinical or behavioral predictions , Then the choice of brain representation may not be important . However , Different brain representations have strong differences in the hypothetical mechanism of psychopathology , Treatment suggestions for potential conflicts . therefore , The best brain representation should be the origin of the disease mechanism ( Not downstream effects ) Theory provides information , And generate verifiable hypotheses for subsequent research . If the purpose is to use only RF MRI Make clinical or behavioral predictions , This lack of biological explanation is acceptable . However , Unfortunately , The explanation of this prediction is often groundless .

chart 3 Examples of differences 4. Suggestions and future development of brain representation To distinguish brain performance and determine the best performance for the test , There are many factors to consider . 1） It's solved “ The challenge of brain representation ” The brain representation of some of the challenges presented in should be preferred over the brain representation that does not address these challenges . 2） It is important to , The size of the brain unit matches the hypothesis . 3） The population used for brain unit definition should match the population of interest . 4） We need to consider the interaction between brain units and summative measurements . We propose a series of validation phases for current and future brain representations . With the development of new brain representations , We propose clear steps (i) Measure the differences between participants , And ideally estimate individual brain units ,(ii) Specifically test the prevalence of different diseases and lifespan populations and different scanners ,(iii) Make a systematic and extensive comparison of the existing brain representation types , Final ,(iv) Explain the brain representation based on multimodal experiments . 5. Conclusion In the field of magnetic resonance imaging, there are differences because of the segmented brain representation . Although the selected brain representation is important for the results and interpretation , However, few articles give clear reasons for adopting specific representations . contrary , Laboratories usually use a specific method , And apply it to all research projects , And relatively little consideration was given to the implicit assumptions of the brain representation they chose . This trend may produce problems that are inconsistent with the basic principles of cumulative science , A research shaft consisting of separate reasoning and assumptions . To break through the limitations of these research areas , Successful collaborative brain mapping and interpretable biomarker discovery , You need to better understand the relationship between different representations of the same data . Once we get a clearer understanding of the different brains and rfMRI Representation of the relationship between data and underlying neurophysiology , Some key concepts , explain , Definitions and terms may need to be redefined or updated . This will require commitment from members of the field and a willingness to test 、 Challenge and modify our assumptions and core principles . Improve rfMRI The interpretability of brain representation will improve the repeatability of results in different research laboratories , And improve rfMRI The real clinical impact of , Provide information for diagnosis and treatment . The guidelines and recommendations presented in this article are intended to bring the wider community together , Set new standards for this field .