QuNex-An integrative platform for reproducible neuroimaging analytics

Abstract

Introduction: Neuroimaging technology has experienced explosive growth and transformed the study of neural mechanisms across health and disease. However, given the diversity of sophisticated tools for handling neuroimaging data, the field faces challenges in method integration, particularly across multiple modalities and species. Specifically, researchers often have to rely on siloed approaches which limit reproducibility, with idiosyncratic data organization and limited software interoperability.

Methods: To address these challenges, we have developed Quantitative Neuroimaging Environment & Toolbox (QuNex), a platform for consistent end-to-end processing and analytics. QuNex provides several novel functionalities for neuroimaging analyses, including a "turnkey" command for the reproducible deployment of custom workflows, from onboarding raw data to generating analytic features.

Results: The platform enables interoperable integration of multi-modal, community-developed neuroimaging software through an extension framework with a software development kit (SDK) for seamless integration of community tools. Critically, it supports high-throughput, parallel processing in high-performance compute environments, either locally or in the cloud. Notably, QuNex has successfully processed over 10,000 scans across neuroimaging consortia, including multiple clinical datasets. Moreover, QuNex enables integration of human and non-human workflows via a cohesive translational platform.

Discussion: Collectively, this effort stands to significantly impact neuroimaging method integration across acquisition approaches, pipelines, datasets, computational environments, and species. Building on this platform will enable more rapid, scalable, and reproducible impact of neuroimaging technology across health and disease.

Keywords: cloud integration; containerization; data processing; diffusion MRI; functional MRI; high-performance computing; multi-modal analyses; neuroimaging.

Figure 1

QuNex provides an integrated, versatile, and flexible neuroimaging platform. (A) QuNex supports processing of input data from multiple species, including human, macaque, and mouse. (B) Additionally, data can be onboarded from a variety of popular formats, including neuroimaging data in DICOM, PAR/REC, NIfTI formats, a full BIDS dataset, or behavioral data from task performance or symptom assessments. (C) The QuNex platform is available as a container for ease of distribution, portability, and execution. The QuNex container can be accessed via the command line and contains all the necessary packages, libraries, and dependencies needed for running processing and analytic functions. (D) QuNex is designed to be easily scalable to accommodate a variety of datasets and job sizes. From a user access point (i.e., the user's local machine), QuNex can be deployed locally, on cloud servers, or via job schedulers in supercomputer environments. (E) QuNex outputs multi-modal features at the single subject and group levels. Supported features that can be extracted from individual subjects include structural features from T1w, T2w, and dMRI (such as myelin, cortical thickness, sulcal depth, and curvature) and functional features from BOLD imaging (such as functional connectivity matrices). Additional modalities, e.g., receptor occupancy from PET (positron emission tomography), are also being developed (see Section Discussion). Features can be extracted at the dense, parcel, or network levels. (F) Importantly, QuNex also provides a comprehensive set of tools for community contribution, engagement, and support. A Software Development Kit (SDK) and GitFlow-powered DevOps framework is provided for community-developed extensions. A forum (https://forum.qunex.yale.edu) is available for users to engage with the QuNex developer team to ask questions, report bugs, and/or provide feedback.

Figure 2

QuNex turnkey functionality and batch engine for high-throughput processing. (A) QuNex provides a “turnkey” engine which enables fully automated deployment of entire pipelines on neuroimaging data via a single command (qunex run_turnkey). An example of a typical workflow with key steps supported by the turnkey engine is highlighted, along with the example command specification. QuNex supports state-of-the-art preprocessing tools from the neuroimaging community (e.g., the HCP MPP; Glasser et al., 2013). For a detailed visual schematic of QuNex steps and commands (see Supplementary Figure 1). (B) The QuNex batch specification is designed to enable flexible and comprehensive “filtering” and selection of specific data subsets to process. The filtering criteria can be specified at multiple levels, such as devices (e.g., Siemens, GE, or Philips MRI scanners), institutions (e.g., scanning sites), groups (e.g., patient vs. controls), subjects, sessions (e.g., time points in a longitudinal study), modalities (e.g., T1w, T2w, BOLD, diffusion), or scan tags (e.g., name of scan). (C) QuNex natively supports job scheduling via LSF, SLURM, or PBS schedulers and can be easily deployed in HPC systems to handle high-throughput, parallel processing of large neuroimaging datasets. The scheduling options enable precise specification of paralellization both across sessions and within session (e.g., parallel processing of BOLD images) for optimal performance and utilization of cluster resources.

Figure 3

Consistent processing at scale and standardized outputs through batch specification. (A) The batch specification mechanism in QuNex is designed to support data processing from single-site and multi-site datasets to produce standardized outputs. Acquisition parameters can be flexibly specified for each sequence. Here, example datasets I (single-site study) and II (multi-site study) illustrate possible use cases, with the sequences in each dataset shown in green text. Although Dataset I does not include T2w scans, and Dataset II contains data from different scanners, all these data can be consistently preprocessed in all modalities to produce standardized output neural features. (B) Parameters can be tailored for each study in the header of the batch processing file. An example is shown with parameters in green text tailored to Site B in Dataset II (similar to those used in HCP datasets; Glasser et al., 2013). Detailed instructions and examples for setting up the batch parameter header for a user's specific study is available in the documentation. (C) QuNex has been highly successful in preprocessing data from numerous publicly available as well as private datasets, totalling over 10,000 independent scan sessions from over 50 different scanners. In some cases, advanced user options can be used to rescue sessions which failed with “out-of-the-box” default preprocessing options. These options include using custom brain masks, control points, or expert file options in Freesurfer (Fischl, ; McCarthy et al., 2015) (see Supplementary material). The number of successful/total sessions is reported in each bar. The number of sessions rescued with advanced options is shown in parentheses, when applicable. The total proportion of successfully preprocessed sessions from each study (including any sessions rerun with advanced options) as well as the grand total across all studies is shown above the bar plots. The majority of the sessions which failed were due to excessive motion in the structural T1w image, which can cause issues with the registration and segmentation. (D) QuNex has been successfully used to preprocess data with a wide range of parameters and from diverse datasets. (Left) QuNex has been tested on MRI data acquired with the three major scanner manufacturers (Philips, GE, and Siemens). Here N_S specifies the number of individual scan sessions that were acquired with each type of scanner. (Middle) QuNex is capable of processing images acquired both with and without simultaneous multi-slice (SMS) acquisition (also known as multi-band acquisition, i.e., Simultaneous Multi-Slice in Siemens scanners; Hyperband in GE scanners; and Multi-Band SENSE in Philips scanners; Kozak et al., 2020). (Right) QuNex has been tested on data from clinical, pharmacology, longitudinal, and basic population-based datasets. Here, N_D specifies the number of datasets; N_S specifies the total number of individual scan sessions in those datasets.

Figure 4

Extracting multi-modal processing features at multiple levels of resolution. Output features from multiple modalities are shown, as an example of a cross-modal analysis that may be done for a study. Here, features were computed from a cohort of N = 339 unrelated subjects from the HCP Young Adult cohort (Van Essen et al., 2013). In addition to cross-modality support, QuNex offers feature extraction at “dense” (i.e., full-resolution), parcel-level and network-level resolutions. All features are shown below at all three resolutions. We used the Cole-Anticevic Brainwide Network Parcellation (CAB-NP) (Glasser et al., ; Ji et al., 2019b), computed using resting-state functional connectivity from the same cohort and validated and characterized extensively in Ji et al. (2019b). (A) Myelin maps, estimated using the ratio of T1w/T2w images (Glasser and Van Essen, 2011). (B) Left arcuate fasciculus computed via diffusion tractography (Warrington et al., 2020). Surface views show the cortical tract termination (white-gray matter boundary endpoints) and volume views show the maximal intensity projection. (C) Structural connectivity of Broca's area (parcel corresponding to Brodmann's Area [BA] 44, green star) (Glasser et al., 2016a). (D) Resting-state functional connectivity of Broca's area (green star). For parcel- and network-level maps, resting-state data were first parcelated before computing connectivity. (E) Task activation maps for the “Story vs. Math” contrast in a language processing task (Barch et al., 2013). For parcel- and network-level maps, task fMRI data were first parcellated before model fitting. (F) (Left) Whole-brain Language network from the CAB-NP (Ji et al., 2019b). (Right) The mean t-statistic within Language network regions from the “Story vs. Math” contrast [shown in (E)] improves when data are first parcellated at the parcel-level relative to dense-level data and shows the greatest improvement when data are first parcellated at the network-level. Error bars show the standard error. (G) (Left) T-statistics computed on the average parcel beta estimates are higher compared to the average T-statistics computed over dense estimates of the same parcel. Teal dots represent 718 parcels from the CAB-NP × 3 Language task contrasts (“Story vs. Baseline”; “Math vs. Baseline”; “Story vs. Math”). (Right) Similarly, T-statistics computed on beta estimates for the network are higher than the average of T-statistics computed across parcels within each network.

Figure 5

General Linear Model (GLM) for single-session modeling of time-series modalities and integrated interoperability with PALM for group-level analytics. (A) The QuNex GLM framework enables denoising and/or event modeling of resting-state and task BOLD images at the individual-session level in a single step. A use case is shown for resting-state BOLD data. At the single-subject level, FL FLthe user can choose to specify FL individual nuisance regressors (such as white matter and ventricular signal and motion parameters) such that they are regressed out of the BOLD timeseries with the qunex preprocess_conc function. The regressors can be per-frame (as shown), per-trial, or even per-block. The GLM outputs a residual timeseries of “denoised” resting-state data as well as one coefficient map per nuisance regressor. The resting-state data for each subject can then be used to calculate subject-specific feature maps, such as seed-based functional connectivity maps with qunex fc_compute_seedmaps. (B) The GLM engine can also be used for complex modeling and analysis of task events, following a similar framework. Event modeling is specified in qunex preprocess_conc by providing the associated event file; the method of modeling can be either assumed (using a hemodynamic response function [HRF] of the user's choosing, e.g. Boynton) or unassumed. Here, an example from the HCP's Language task is shown. The two events, “Story” and “Math,” are convolved with the Boynton HRF to build the subject-level GLM. As with the resting-state use case shown in (A), the GLM outputs the single-subject residual timeseries (in this case “pseudo-resting state”) as well as the coefficient maps for each regressor, here the Story and Math tasks. (C) Connectivity maps from all subjects can then be entered into a group-level GLM analysis. In this example, the linear relationship between connectivity from the primary somatosensory area (S1) seed and age across subjects is tested in a simple GLM design with one group and one explanatory variable (EV) covariate, demeaned age. QuNex supports flexible group-level GLM analyses with non-parametric tests via Permutation Analysis of Linear Models (PALM, Winkler et al., 2014), through the qunex run_palm function. The specification of the GLM and individual contrasts is completely configurable and allows for flexible and specific hypothesis testing. Group-level outputs include full uncorrected statistical maps for each specified contrast as well as p-value maps that can be used for thresholding. Significance for group-level statistical maps can be assessed with the native PALM support for TFCE (Winkler et al., , shown) or cluster statistics with familywise error protection (FWEP). (D) The subject-level task coefficient maps can then be input into the qunex run_palm command along with the group-level design matrix and contrasts. The group-level output maps show the differences in activation between the Story and Math conditions. (E) QuNex also supports multi-variate and joint inference tests for testing hypotheses using data from multiple modalities, such as BOLD signal and dMRI. Example connectivity matrices are shown for these two modalities, with the S1 seed highlighted. Similar to the use cases shown above, maps from all subjects can be entered into a group-level analysis with a group-level design matrix and contrasts using the qunex run_palm command. In this example, the relationship between age and S1-seeded functional connectivity and structural connectivity is assessed using a Hotelling's T² test and Fisher's X². The resulting output maps show the unthresholded and thresholded (p < 0.05 FWEP, 10,000 permutations) relationship between age and both neural modalities.

Figure 6

QuNex enables neuroimaging workflows across different species. (A) Structural features for exemplar macaque and human data, including surface reconstructions and segmentation from FreeSurfer. Lower panel shows output myelin (T1w/T2w) maps. (B) Functional features for exemplar macaque and human showing BOLD signal mapped to both volume and surface. Lower panels show and resting-state functional connectivity seeded from the lateral geniculate nucleus of the thalamus (green arrow). (C) Diffusion features for exemplar macaque and human data, showing whole-brain fractional anistropy, and volume and surface terminations of the left optic radiation tract. Lower panels show the structural connectivity maps seeded from the lateral geniculate nucleus of the thalamus (green arrow). Gray scale reference bars in each panel are scaled to 25 mm.

Figure 7

Features included in QuNex and comparisons to other neuroimaging software. A list of tools integrated into QuNex along with supported QuNex functionalities. We also list functionalities and tools currently available in some popular neuroimaging pipelines/environments, including fMRIPrep (Esteban et al., 2019), QSIPrep (Cieslak et al., 2021), HCP (Glasser et al., 2013), UK Biobank (Alfaro-Almagro et al., 2018), nipype (Gorgolewski et al., 2016), micapipe (Cruces et al., 2022), FuNP (Park et al., 2019), NeuroDebian (Halchenko and Hanke, 2012), BrainVoyager (Goebel, 2012), and LONI (Dinov et al., 2009).