香京julia种子在线播放

Technological constraints, such as storage space, download speed, and processing power, play a significant role in the feasibility of performing collaborative analyses on large datasets (Homer et al., 2008; McGuire et al., 2011) such as neuroimaging data. Existing data repositories can contain high-resolution neuroimaging files covering tens of thousands of subjects, with sizes ranging from megabytes to multiple petabytes. Downloading the MPI-Leipzig Mind-Brain-Body dataset (Babayan et al., 2022) (369.78 GB) at the global median download speed of 76.32 Mbps² onto a modern MacBook Pro with 512 GB of storage³ would take 11 h and 33 min, consuming 72.2 percent of the machine's total storage space. The requirements for storage space and download time can increase when an analysis involves aggregating multiple large datasets. Additionally, processing power may be a limiting factor for performing computations, particularly when certain types of analyses are designed to run on specific hardware like GPUs, which can demand resources beyond the capacity of smaller research groups or institutions with limited budgets.

Due to the potentially sensitive nature of neuroimaging datasets, their use in collaborative analysis is often restricted by policies intended to preserve privacy. Collaboration methods include aggregating data in a centralized repository or using Data Usage Agreements (DUAs) (Thompson et al., 2014, 2017). These methods can be cumbersome and, in some cases, insufficient. DUAs may take months or even years to approve without any guarantee of the data's utility. Data sharing might be limited by law, policy, or proprietary restrictions, largely driven by re-identification concerns. In situations where only summary data can be shared, differences in analysis methodology may result in inconsistent measures for meta-analysis (Rootes-Murdy et al., 2022).

Administrative challenges can arise when collaborating on an analysis, as various steps demand researchers' time and attention. These steps may include communicating between agencies, formulating and signing data-sharing agreements, agreeing on data preparation and analysis processes, procuring technical resources, monitoring and auditing processes, performing data transfer, initiating computations, disseminating results of analyses, and so on.

The efficiency of collaborative analysis is influenced by how quickly these manual steps are executed. Synchronized availability of researchers can present a barrier to the collaboration process. When researchers work asynchronously, each step in a serial process requiring manual interaction introduces potential delays. This can be particularly challenging when researchers are distributed across multiple time zones or have limited time to perform manual tasks. Furthermore, researchers' availability may be constrained by the need for expertise and authority, such as having the authority to sign a data-sharing agreement or the technical expertise to run the appropriate Python script against a dataset. Often, these manual steps must be executed for each new analysis, which can slow down and even impede collaborative analysis. By addressing these administrative barriers, research teams can more effectively collaborate and streamline their analysis processes, ultimately contributing to the advancement of neuroimaging research.

Variability in methodological approaches to data processing and analysis can make reproducing studies challenging (Vogt, 2023). To validate results, researchers must adhere to the exact methodology used in the original study, which necessitates clear communication of the specific methods employed. However, as methods are often chosen on a case-by-case basis, replicating studies can be time-consuming and difficult (Esteban et al., 2019), and sometimes even impossible. Moreover, when multiple studies adopt different methodologies, combining their results meaningfully becomes challenging, hindering the execution of meta-analyses.

To overcome these barriers, we introduce COINSTAC,⁴ a tool that supports federated analysis for neuroimaging data.

Federated analysis (also federated learning, or decentralized analysis) (Plis et al., 2016; Kairouz et al., 2021; Rootes-Murdy et al., 2022) allows for multiple datasets to be used in analyses without source data being directly shared. Instead, data holders run computations on their local data and only share the outputs, which are often group-level data derivatives or summary statistics. For example, sites may compute an average or other summary on their local data and send that information. Typically, these summaries are much smaller, meaning that the source data are not shared, thereby removing the technical challenges associated with dataset transfer. A second potential benefit is additional privacy guarantees for the data holders. From a purely policy perspective, datasets are analyzed without being moved from their original location and data holders can determine which computations are and are not allowed on their data. From a technical perspective, strong end-to-end encryption can prevent third parties from acquiring the data derivatives. Depending on the trust model, additional privacy protections are possible, including emerging technologies like secure multiparty computation and differential privacy (Dwork and Roth, 2013; Bonawitz et al., 2016, 2017; Heikkilä et al., 2020; Imtiaz et al., 2021; Senanayake et al., 2022).

The Collaborative Informatics and Neuroimaging Suite Toolkit for Anonymous Computation (COINSTAC) (see text footnote 4) (Plis et al., 2016; Ming et al., 2017; Gazula et al., 2020, 2023; Turner et al., 2022) is a tool developed to support federated analysis specifically for neuroimaging data by overcoming the aforementioned barriers to collaboration through the use of federated analysis and standardization of collaboration methods. COINSTAC enables researchers to run decentralized neuroimaging analyses to perform larger collaborative studies (Rootes-Murdy et al., 2022; Turner et al., 2022). As of now, COINSTAC has attracted 115 users and has been downloaded 2,386 times, showcasing its growing reach and impact within the research community.

The COINSTAC desktop application provides an easy-to-use graphical user interface (GUI) for coordinating and executing federated analysis pipelines among multiple collaborators. Image preprocessing and a variety of univariate and multivariate approaches (e.g., VBM regression, group ICA) can be completed within the app.

For a comprehensive understanding of COINSTAC, its functionalities, and usage, readers are encouraged to refer to the following papers (Plis et al., 2016; Ming et al., 2017; Gazula et al., 2020, 2023; Turner et al., 2022).

One limitation of the original implementation of COINSTAC is that it requires synchronized coordination (Jwa and Poldrack, 2022), users have to coordinate among data owners to confirm their systems are online, that the data are organized within the same structure and that the data are mapped properly within the COINSTAC system. The need for a centralized coordinator can delay contingent analyses.

In this paper, we address this limitation by showcasing a method for hosting both private and public datasets where the datasets are persistently accessible for analysis using COINSTAC without the need for synchronized effort from data owners. Analysis of public datasets is made more accessible by removing the need to find, download, preprocess, and prepare datasets for analysis. We provide curated data vaults for various openly available neuroimaging data which COINSTAC users can simply include in their analyses. Access to private datasets can be restricted to a list of computations approved by the vault owner. Standardizing access to data vaults in the COINSTAC system simplifies analysis, optimizes computational performance, and promotes the reusability of neuroimaging datasets.

To understand how Vaults improve the workflow of federated analysis in COINSTAC, we will describe the COINSTAC system and how it is used.

The main components of the COINSTAC system are: the desktop application, the central server, and computation containers. The desktop application provides a graphical user interface (GUI) and manages local computation containers used to participate in federated analyses. The central server manages the central database and runs the containers that act as the inner node in federated analyses.

In the COINSTAC desktop application, users join collections of users called “consortia” to collaborate on an analysis pipeline. A consortium is a group formed by individual COINSTAC users, each with their machine that is capable of being a node in a federated analysis pipeline. Each member within a consortium will act as a node in the federated analysis group by running local computations inside of a container on their system.

The following is how a researcher would use the COINSTAC user interface to create a consortium and run a federated analysis pipeline:

The TReNDS VBM COBRE Vault contains structural MRI images from 152 participants, approximately half healthy volunteers and half individuals diagnosed with schizophrenia, collected as part of the Mind Research Network COBRE study (Aine et al., 2017). The Vault includes gray matter MRI images that have been run through a VBM preprocessing pipeline in the SPM toolbox. In addition, we have demographic information, symptom severity scales, and cognitive measures to select from when building a desired model. Figure 5 shows the beta images from running VBM regression on all the voxels from normalized smoothed gray matter images from the TReNDS COBRE Vault. Age, sex, and diagnosis information were used as covariates in the regression model. Results show decreases in brain volume with age, reduced volume in visual areas and along the gray/white boundary in females, and reduced volume in insular-temporal and medial frontal regions in schizophrenia patients, consistent with previous results.

The following section describes this use-case with 55 participant's structural MRI scans collected under MCIC project (Gollub, 2013). The results from running regression on the normalized smoothed gray matter images from this project are shown in Figure 6.

Using the MCIC dataset, we similarly see widespread reduction in brain volume for age, visual and gray/white boundary reductions in volume in females, and insular-temporal and medial frontal (as well as more wide spread) reductions in schizophrenia patients.

The TReNDS VBM COBRE Vault was combined with the MCIC dataset, allowing for an increased sample size, in the same regression analysis to examine diagnostic effects while accounting for age and sex. The combined dataset was largely consistent with the individual site analysis, with the exception of the male/female effect which shows a more complex pattern of increases and decreases, though still largely conforming to reductions in white/gray matter boundary and primary visual area volumes (Gupta et al., 2015). Results of this study are shown in Figure 7.

This Vault contains data from 152 subjects, approximately half controls and half individuals with chronic schizophrenia, collected as part of the Mind Research Network COBRE study.⁶ The Vault includes cortical and sub-cortical volumetric and surface-based measurements from two FreeSurfer atlases, Desikan-Killiany and Destrieux. In addition, we have a total of 11 variables across demographic, cognitive, and substance use to select from when building a desired model.

We ran Ridge regression on the above Vault data on Freesurfer volumetric and surface based measurements on about 500 regions of interest. We noticed the following differences between controls and patients.

Controls have higher values in temporal lobe, as shown in the thickness measurements of tables (Tables 1–5).

This Vault contains data from 922 children and adolescents (ages 6–22, 603 Male and 319 female), collected as part of the Healthy Brain Network study (Alexander et al., 2017). The Vault includes gray matter segmentation data from an SPM VBM preprocessing pipeline. In addition, we have a total of 11 variables across various demographic, cognitive and substance use domains to select from when building a desired model.

Figure 8 shows the beta images from running regression on all the voxels from normalized smoothed gray matter images from the CMI VBM VAULT. Age and sex were used as covariates in the regression model. Results were largely consistent with those from the MCIC and COBRE analyses, showing widespread volume reductions with age, and reductions along the gray/white matter boundary in females.

Group ICA (Calhoun et al., 2001) is one of the frequently used preprocessing computations for neuroimaging data. Data preprocessed with group ICA can be used to perform different types of analyses. This GICA Vault comprises data from 189 subjects from the COBRE project analyzed with Neuromark template which uses 66 predefined ROIs. This Vault data includes independent component analysis (ICA) maps, Functional network connectivity maps (FNC) data etc. that have been generated using spatially constrained ICA with the Neuromark_fMRI_1.0 template (available in the GIFT software)⁷,⁸ including 53 intrinsic networks (components). This Vault data can be readily used for secondary analysis like mancova. In this case, we use the GICA pre-processed data from the Vault to perform univariate regression analysis, the results of which are shown in Figure 9.

The Neuromark fMRI domains identified in Du et al. Briefly, these seven identified network templates were divided based on anatomical and functional properties (Du et al., 2020). In each subfigures, one color in the composite maps corresponds to an intrinsic connectivity network (ICN). The Neuromark_fMRI_1.0 template is available in the GIFT software (Figure 10).

COINSTAC Vaults offer numerous benefits, but there are also limitations and challenges to consider, particularly in the areas of data privacy and security, and resource usage.

One concern is that allowing arbitrary summary queries on a dataset might enable an attacker to reconstruct the data. To mitigate such risks, the system must be privacy-preserving from “end-to-end,” incorporating techniques like secure multiparty computation or differential privacy. Implementing these methods can be difficult due to floating point implementation issues (Mironov, 2012; Ilvento, 2020a,b) and the introduction of noise, which may increase error or variance in the analysis results.

While differentially private algorithms can provide stronger privacy guarantees, sharing data derivatives without differential privacy might be adequate in some situations, depending on the trust model and privacy concerns of data holders. These issues should be addressed on a case-by-case basis.

Vault owners can currently restrict computations on their data to a pre-approved list. To enhance privacy protection, further improvements are recommended. Potential solutions include allowing Vault owners to:

Another challenge is handling slowdowns or crashes during resource-intensive analyses due to high compute usage. To address this issue, Vault owners can be given more control over resource usage and compute capacity. They could limit the number of concurrent computations and overall CPU usage. Improving compute capacity could involve strategies like deploying multiple instances behind a load balancer or dynamically scaling resources.

Additional challenges include data distribution, network bandwidth, and communication speed. Federated learning and open-source solutions can help address some of these problems, but further research and development are needed to optimize COINSTAC Vaults' performance in various research settings. Our “Decentralized Sparse Deep Artificial Neural Networks in COINSTAC (CPU and GPU enabled)” algorithm allows users to save network bandwidth when transferring thousands of derived data/machine learning parameters across nodes.

In summary, COINSTAC Vaults mark a significant advancement in federated neuroimaging research, data privacy preservation, and open science promotion. By tackling the existing limitations and challenges, COINSTAC Vaults can further improve collaboration and innovation within the field.

The long-term vision for COINSTAC and COINSTAC Vaults includes: