Canonical and Replicable Multi-Scale Intrinsic Connectivity Networks in 100k+ Resting-State fMRI Datasets

Armin Iraji, Zening Fu, Ashkan Faghiri, Marlena Duda, J. Chen, Srinivas Rachakonda, Thomas P. DeRamus, Peter Kochunov, Bhim M. Adhikari, Ayşenil Belger, Judith M. Ford, Daniel H. Mathalon, Godfrey D. Pearlson, Steven G. Potkin, Adrian Preda, Jessica A. Turner, Theo G.m. van Erp, Juan Bustillo, Kaihua Yang, K Ishizuka, Akira Sawa, Kent E. Hutchison, Elizabeth Osuch, Jean Théberge, Chris Abbott, Bryon A. Mueller, Dongmei Zhi, Chuanjun Zhuo, Sixing Liu, Yong Xu, Mustafa S. Salman, Jiahang Liu, Yuhui Du, Jing Sui, Tülay Adalı, Vince D. CalhounbioRxiv 2022 · DOI

TL;DR

This paper tackles the challenge of reliably identifying consistent brain functional networks (Intrinsic Connectivity Networks or ICNs) across diverse individuals and datasets using resting-state fMRI. They achieve this by developing a robust, multi-scale ICN template from over 100,000 individuals and then demonstrating how to accurately estimate subject-specific ICNs using this template, paving the way for more standardized and clinically applicable brain network analysis.

What Problem Does This Paper Address?

Resting-state fMRI (rsfMRI) is a powerful tool for understanding brain function, but its clinical utility is hampered by the difficulty in obtaining consistent and comparable functional network estimates across different individuals, datasets, and even within the same individual over time. This variability makes it hard to identify true brain alterations in disorders or to apply findings to single subjects. Existing methods often struggle with individual differences, data quality variations, and the inherent "noise" in rsfMRI signals, limiting the ability to create a universal "map" of brain functional networks. The authors aim to overcome this by developing a reliable, large-scale ICN template and a method for estimating subject-specific ICNs that maintain cross-subject correspondence.

Methods

The core of their approach involves two main stages:

  1. ICN Template Generation:
    • Data Aggregation and Quality Control: They combined rsfMRI data from over 100,000 individuals, meticulously filtering it down to ~58,000 subjects that met stringent quality control (QC) criteria (e.g., minimum scan length, low motion).
  2. Multi-Model-Order Independent Component Analysis (ICA): They employed a sophisticated ICA technique called group-level multi-model-order spatial ICA (gr-msICA). This method allows for the identification of ICNs at various spatial scales simultaneously, without imposing artificial boundaries. They ran this on subsamples of their QC-passed data multiple times to ensure stability and then selected the most distinct ICNs to form their final template.
  3. Replicability Check: Crucially, they validated their template by applying the same analysis to the QC-failed dataset, demonstrating that their template is robust to variations in data quality.
  4. Subject-Level ICN Estimation:
    • Group-Informed Network Estimation: They used a multivariate-spatially constrained ICA approach (specifically, MOO-ICAR) that leverages the generated ICN template to guide the estimation of ICNs for individual subjects. This "group-informed" approach helps to constrain the analysis and improve accuracy, especially with noisy or shorter datasets.
  5. Evaluation Criteria: They introduced two novel criteria to systematically evaluate the success of subject-level ICN estimation, focusing on whether the estimated ICNs provide unique, subject-specific information beyond predefined anatomical regions and are distinct from other ICNs.

Datasets

  • Primary Dataset: Over 100,000 rsfMRI subjects from private and public datasets.
  • QC-Passed Dataset: Approximately 57,709 individuals after quality control.
  • QC-Failed Dataset: Approximately 42,808 individuals who did not pass QC, used for replicability testing.
  • Evaluation Datasets: Functional Imaging Biomedical Informatics Research Network (FBIRN) and Human Connectome Project (HCP) datasets were used to evaluate subject-level estimation under different data specifications (e.g., resolution, length).

Key Insights and Ideas

  • The Power of Scale: Aggregating data from over 100,000 individuals is crucial for creating a truly representative and robust ICN template that captures the diversity of brain functional organization.
  • Multi-Scale ICNs are Key: Brain networks exist at different spatial scales. Using multi-model-order ICA allows for the simultaneous identification of both large-scale and fine-grained networks, providing a more complete picture of brain function.
  • Template-Guided Estimation is Essential: Relying solely on individual subject data can lead to unstable and incomparable results. Using a well-established template significantly improves the accuracy and reliability of subject-specific ICN estimation.
  • Data Quality Matters, But Robustness is Achievable: While data quality is important, the authors show that their template can be reliably identified even in data that doesn't meet strict QC criteria, highlighting the robustness of their approach.
  • Data Length Trade-offs: Longer scan times don't always equate to better results. Increasing data length can sometimes reduce an ICN's subject-level specificity, suggesting a need to optimize scan duration rather than simply maximizing it.
  • Spatial Smoothing is a Double-Edged Sword: While smoothing can improve similarity and reliability, it can also mask important individual differences and dynamic brain activity. Careful consideration of smoothing is needed for interpretation.

Key Results

  • A Reliable 105-ICN Template: They successfully generated a template consisting of 105 distinct, multi-spatial-scale ICNs. This template was highly replicable, showing high spatial similarity (>0.8) even with an unseen QC-failed dataset.
  • Subject-Specific ICNs are Feasible: Using their template and MOO-ICAR, they demonstrated the ability to estimate subject-specific ICNs that are statistically similar to the template.
  • Data Length and Resolution Impact Estimation: The success of subject-level ICN estimation is influenced by data length and spatial resolution. For instance, the HCP dataset (with longer scans) allowed for the estimation of all 105 ICNs, while the FBIRN dataset (with shorter scans and lower resolution) had some ICNs that did not meet the criteria for subject-specific information.
  • Larger ICNs are More Stable: Larger-scale ICNs tend to be more stable and require less data for reliable subject-level estimation compared to smaller, more granular ICNs.
  • Subject Specificity Varies by Network: Certain ICNs, particularly those associated with frontoparietal networks, exhibit higher subject-specificity (i.e., better at distinguishing individuals) than those in subcortical or cerebellar regions.

Limitations and Future Work

  • Data Length Optimization: The study highlights that longer scans don't always improve subject-level specificity, suggesting future work should focus on optimizing data length rather than simply maximizing it.
  • Spatial Smoothing Control: The impact of spatial smoothing on results needs to be carefully controlled for when comparing findings across studies or data lengths.
  • Brain Dynamics: The study acknowledges that incorporating and assessing brain dynamics (i.e., how ICNs change over time) was not explored and is a crucial area for future research.
  • Developing Advanced Techniques: The authors plan to develop more advanced group-informed network estimation techniques, explore higher model order ICA with smaller step sizes, and release their Neuromark templates for other imaging modalities.
  • Understudied Evaluation Criteria: The paper calls for further investigation into the two novel criteria they introduced for evaluating subject-level ICN estimation, an area they note has been overlooked.

Summary updated

Abstract

Abstract Resting-state functional magnetic resonance imaging (rsfMRI) has shown considerable promise for improving our understanding of brain function and characterizing various mental and cognitive states in the healthy and disordered brain. However, the lack of accurate and precise estimations of comparable functional patterns across datasets, individuals, and ever-changing brain states in a way that captures both individual variation and inter-subject correspondence limits the clinical utility of rsfMRI and its application to single-subject analyses. We posit that using reliable network templates and advanced group-informed network estimation approaches to accurately and precisely obtain individualized (dynamic) networks that retain cross-subject correspondence while maintaining subject-specific information is one potential solution to overcome the aforementioned barrier when considering cross-study comparability, independence of subject-level estimates, the limited data available in single studies, and the low signal-to-noise ratio (SNR) of rsfMRI. Toward this goal, we first obtained a reliable and replicable network template. We combined rsfMRI data of over 100k individuals across private and public datasets and selected around 58k that meet quality control (QC) criteria. We then applied multi-model-order independent component analysis (ICA) and subsampling to obtain reliable canonical intrinsic connectivity networks (ICNs) across multiple spatial scales. The selected ICNs (i.e., network templates) were also successfully replicated by independently analyzing the data that did not pass the QC criteria, highlighting the robustness of our adaptive template to data quality. We next studied the feasibility of estimating the corresponding subject-specific ICNs using a multivariate-spatially constrained ICA as an example of group-informed network estimation approaches. The results highlight that several factors, including ICNs themselves, data length, and spatial resolution, play key roles in successfully estimating the ICNs at the subject level. Large-scale ICNs, in general, require less data to achieve a specific level of spatial similarity with their templates (as well as within- and between-subject spatial similarity). Moreover, increasing data length can reduce an ICN’s subject-level specificity, suggesting longer scans might not always be desirable. We also show spatial smoothing can alter results, and the positive linear relationship we observed between data length and spatial smoothness (we posit that it is at least partially due to averaging over intrinsic dynamics or individual variation) indicates the importance of considering this factor in studies such as those focused on optimizing data length. Finally, the consistency in the spatial similarity between ICNs estimated using the full-length of data and subset of it across different data lengths may suggest that the lower within-subject spatial similarity in shorter data lengths is not necessarily only defined by lower reliability in ICN estimates; rather, it can also be an indication of brain dynamics (i.e., different subsets of data may reflect different ICN dynamics), and as we increase the data length, the result approaches the average (also known as static) ICN pattern, and therefore loses its distinctiveness.
BibTeX
@misc{iraji2022canonical,
  author = {Armin Iraji and Zening Fu and Ashkan Faghiri and Marlena Duda and J. Chen and Srinivas Rachakonda and Thomas P. DeRamus and Peter Kochunov and Bhim M. Adhikari and Ayşenil Belger and Judith M. Ford and Daniel H. Mathalon and Godfrey D. Pearlson and Steven G. Potkin and Adrian Preda and Jessica A. Turner and Theo G.M. van Erp and Juan Bustillo and Kaihua Yang and K Ishizuka and Akira Sawa and Kent E. Hutchison and Elizabeth Osuch and Jean Théberge and Chris Abbott and Bryon A. Mueller and Dongmei Zhi and Chuanjun Zhuo and Sixing Liu and Yong Xu and Mustafa S. Salman and Jiahang Liu and Yuhui Du and Jing Sui and Tülay Adalı and Vince D. Calhoun},
  title = {{Canonical and Replicable Multi-Scale Intrinsic Connectivity Networks in 100k+ Resting-State fMRI Datasets}},
  year = {2022},
  doi = {10.1101/2022.09.03.506487},
  howpublished = {bioRxiv},
}
Figure