Atlas-guided parcellation: Individualized functionally-homogenous parcellation in cerebral cortex

https://doi.org/10.1016/j.compbiomed.2022.106078Get rights and content

Highlights

  • An efficient and flexible individualized parcellation framework of the cerebral cortex without parcel alignment.

  • The framework considers both functional organization distribution of prior atlases and homogeneity of individual parcels.

  • The framework can capture individualized brain topographic features with excellent fingerprinting and disease evaluation.

  • DMN is the most representative, stable, and individual-identifiable network in the resting state based on brain topography.

  • Parkinson’s Disease evaluation indicates the gradual topographic changes of our brain along with the disease progression.

Abstract

Resting-state Magnetic resonance imaging-based parcellation aims to group the voxels/vertices non-invasively based on their connectivity profiles, which has achieved great success in understanding the fundamental organizational principles of the human brain. Given the substantial inter-individual variability, the increasing number of studies focus on individual parcellation. However, current methods perform individual parcellations independently or are based on the group prior, requiring expensive computational costs, precise parcel alignment, and extra group information. In this work, an efficient and flexible parcellation framework of individual cerebral cortex was proposed based on a region growing algorithm by merging the unassigned and neighbor vertex with the highest-correlated parcel iteratively. It considered both consistency with prior atlases and individualized functional homogeneity of parcels, which can be applied to a single individual without parcel alignment and group information. The proposed framework was leveraged to 100 unrelated subjects for functional homogeneity comparison and individual identification, and 186 patients with Parkison's disease for symptom prediction. Results demonstrated our framework outperformed other methods in functional homogeneity, and the generated parcellations provided 100% individual identification accuracy. Moreover, the default mode network (DMN) exhibited higher functional homogeneity, intra-subject parcel reproducibility and fingerprinting accuracy, while the sensorimotor network did the opposite, reflecting that the DMN is the most representative, stable, and individual-identifiable network in the resting state. The correlation analysis showed that the severity of the disease symptoms was related negatively to the similarity of individual parcellation and the atlases of healthy populations. The disease severity can be correctly predicted using machine learning models based on individual topographic features such as parcel similarity and parcel size. In summary, the proposed framework not only significantly improves the functional homogeneity but also captures individualized and disease-related brain topography, serving as a potential tool to explore brain function and disease in the future.

Introduction

Understanding the topographic organization of the human brain and parcellating it into distinct cortical regions is one of the greatest challenges in neuroscience [1], for more than a century [2]. The defined atlases support the studies of brain structure, function, and connectivity, as the basis of general systems-neuroscience research [2,3]. In addition, as prior knowledge, atlases can appropriately divide the brain into several networks consisting of interconnected regions, each associated with a specific function [[3], [4], [5], [6]]. The brain atlases also facilitate computational analysis of magnetic resonance imaging (MRI), which is capable of significantly reducing the data dimension from hundreds of thousands of voxels/vertices to a bunch of regions/nodes [3]. The most frequently used anatomical atlases are generally delineated upon macrostructural characteristics such as sulci and gyri [3,7]. However, such atlases are suggested to be heterogeneous in functional and connectional properties [3], and may not ensure the consistency between the anatomical and functional boundaries [8], challenging the subsequent investigations.

Over the past few decades, MRI technology has made it possible for researchers to non-invasively explore the functional parcellation in the living brain, generating many functional atlases such as Yeo et al. [6], Shen et al. [9], Fan et al. [5], Glasser et al. [4] and Schaefer et al. [10]. These atlases divide the whole brain into several to thousands of functionally homogeneous subregions, which are frequently adopted for region of interest (ROI)/network node definition. Brain atlases established with resting-state functional MRI (rs-fMRI) are widely used in connectional fingerprinting, behavioral prediction, aging and disease exploration, etc [[11], [12], [13], [14]]. The most prevalent functional characterization methods obtain the representative signal of a region by averaging the time series of voxels/vertices within the defined parcel [7,15], which requires an underlying assumption that the time series are similar within this parcel, and inhomogeneous ROIs may cause biased conclusions [15]. Although the aforementioned atlases driven by populations can represent the brain organization of a general population, previous studies have reported substantial inter-individual variability in brain networks and regional topography [3,8,16]. Specifically, despite their similarity, there are significant differences in the size, shape and position of subregions between the individual and group atlases [13]. Direct use of group atlases may be inadequate to accurately characterize individual functions, missing important features or leading to possible misunderstanding. Therefore, individualized atlases are preferable over group-averaged atlases for each participant, especially fingerprinting tasks [13].

Currently, the most popular methods of generating an individual atlas are applying parcellation approaches directly to individual data [17] or performing individual parcellation with group information prior [16,18,19]. The former such as clustering [17] and boundary mapping [20] technology requires expensive computational costs and preset parameters (e.g., the number of parcels and boundary threshold). Importantly, independent parcellation is challenged in aligning labels among subjects [8], which prevents direct comparisons between individuals. Another individualized parcellation mapping from populations and atlases addresses the label alignment by matching with or combining the group information, e.g., constraining by group parcellation [18], minimizing group connectivity variability while maximizing group homogeneity [21], matching with the group connectivity profiles [16], and clustering starting with group cluster centroids [22]. Kong et al. proposed multi-session hierarchical Bayesian models to take into account intra- and inter-subject variability [19] as well as spatial contiguity [23] to generate individual-specific networks and improve the performance of behavioral prediction. Zhao and colleagues leveraged the fast matching approach to achieve the individualized mapping from population-based atlases, which perfectly identifies the different individuals [8]. However, the above methods still require extra group information such as group average connectivity matrix and inter-subject variability, which may not be applicable to a single or a few individuals. Additionally, although the generated atlases can be compared with the prominent atlases by setting the same number of parcels or by performing community detection algorithms such as Infomap [16,24], it's challenging for most individualized parcellation approaches to precisely align each parcel with known atlases.

To address the aforementioned issues, we proposed a simple and efficient method without any group information to transfer the known atlases to individuals. This atlas-guided parcellation (AGP) method utilized well-known atlases to initialize individual parcel seeds, and then performed region growing based on the idea of maintaining maximum functional homogeneity. We hypothesized that the proposed method can produce parcels that were as functionally homogenous as possible while recognizing subject-specific topographic fingerprints. To this end, we estimated the individual specificity of the generated atlases on 100 unrelated participants from the Human Connectome Project (HCP) and compared two widely used measures of functional homogeneity with other methods. Additionally, we expected that our method can also capture disease-related topographic characterization. We applied our method to 186 patients with Parkinson's disease (PD) from the Parkinson Progression Marker Initiative (PPMI) database and predicted the disease severity based on individual topographic features. All results were in line with our expectations, and the 7 prior functional atlases and AGP codes are openly available at https://github.com/ly6ustc/Atlas-guided-parcellation.

Section snippets

Participants and neuroimaging acquisition

The first rs-fMRI dataset included 100 unrelated healthy subjects aged 22–36 years (46 males, 54 females), openly available from the HCP database (https://db.humanconnectome.org) [25]. All subjects acquired two sessions (rfMRI_REST1_LR and rfMRI_REST2_LR) on two different days. Two rs-fMRI images were acquired on a customized Siemens 3-T connectome-Skyra scanner with multiband pulse sequences: time point = 1200, multiband factor = 8, bandwidth (BW) = 2290 Hz/px, repetition time = 720 ms, echo

Individual parcellations based on prior atlases

Based on the proposed AGP method, the left-hemisphere parcellations of three exemplary HCP subjects and corresponding prior atlases were exhibited in Fig. 2, and the right-hemisphere results were provided in Fig. S1. The functional organization distribution of individuals was largely consistent with the prior atlases, but significantly different topography such as parcel size and shape existed in the specific parcels, which is more noticeable with the number of parcels decreasing. For Gordon

Discussion

MRI-based parcellation has strengthened our understanding of the fundamental organizational principles of the human brain, contributing to the exploration of brain function in normal and diseased states. In this study, we proposed a general and efficient framework to generate individual parcellation of the cerebral cortex from known atlases without troublesome parcel alignment. This framework can independently perform individual parcellation without any group information as well as take into

Conclusion

In this paper, a novel and flexible framework is proposed to generate the individual cerebral cortex parcellation based on a known atlas. This framework well balances the population level commonality and subject level variability. It provides an efficient way to obtain the individual parcellation which is capable of aligning with the functional organization of prior atlases and achieving superior homogeneity compared with other methods. Additionally, the generated parcellations can capture

Declaration of competing interest

The authors declare no competing financial interests.

Acknowledgements

This work was supported by the National Natural Science Foundation of China (Grants 82272070, 32271431 and 61922075), Anhui Huami Information Technology Co., Ltd., and the Cyrus Tang Foundation.

References (53)

  • C. Caballero-Gaudes et al.

    Methods for cleaning the BOLD fMRI signal

    Neuroimage

    (2017)
  • G.S. Wig et al.

    An approach for parcellating human cortical areas using resting-state correlations

    Neuroimage

    (2014)
  • T.O. Laumann

    Functional system and areal organization of a highly sampled individual human brain

    Neuron

    (2015)
  • S. Mueller

    Individual variability in functional connectivity architecture of the human brain

    Neuron

    (2013)
  • K.A. Schindlbeck et al.

    Network imaging biomarkers: insights and clinical applications in Parkinson's disease

    Lancet Neurol.

    (2018)
  • B.T.T. Yeo et al.

    A modern map of the human cerebral cortex

    Nature

    (2016)
  • S.B. Eickhoff et al.

    Imaging-based parcellations of the human brain

    Nat. Rev. Neurosci.

    (2018)
  • M.F. Glasser

    A multi-modal parcellation of human cerebral cortex

    Nature

    (2016)
  • L.Z. Fan

    The human brainnetome atlas: a new brain atlas based on connectional architecture

    Cerebr. Cortex

    (2016)
  • B.T.T. Yeo

    The organization of the human cerebral cortex estimated by intrinsic functional connectivity

    J. Neurophysiol.

    (2011)
  • M.L. Stanley

    Defining nodes in complex brain networks

    Front. Comput. Neurosci.

    (2013)
  • J.J. Zhao et al.

    Functional parcellation of individual cerebral cortex based on functional MRI

    Neuroinformatics

    (2020)
  • A. Schaefer

    Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI

    Cerebr. Cortex

    (2018)
  • E.S. Finn

    Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity

    Nat. Neurosci.

    (2015)
  • M.C. Campbell

    Proteinopathy and longitudinal changes in functional connectivity networks in Parkinson disease

    Neurol.

    (2020)
  • J. Bijsterbosch

    Challenges and future directions for representations of functional brain organization

    Nat. Neurosci.

    (2020)
  • Cited by (17)

    • An Overview of Brain Fingerprint Identification Based on Various Neuroimaging Technologies

      2024, IEEE Transactions on Cognitive and Developmental Systems
    View all citing articles on Scopus
    View full text