PyNets: Reproducible Ensemble Graph Analysis of Functional and Structural Connectomes

Presented During:


Poster No:

1967 

Submission Type:

Abstract Submission 

Authors:

Derek Pisner1, Ryan Hammonds2

Institutions:

1University of Texas at Austin, Austin, TX, 2University of Texas at Dallas, Dallas, TX

First Author:

Derek Pisner  
University of Texas at Austin
Austin, TX

Co-Author:

Ryan Hammonds  
University of Texas at Dallas
Dallas, TX

Introduction:

Connectomics remains a nascent subfield of neuroscience with a constantly evolving set of methods. Although connectomes may afford us the ability to study fine-grained, high-dimensional individual differences, that gain would seem to come at additional costs to reproducibility (that we cannot afford). More specifically, in estimating a connectome from neuroimaging data, the researcher is forced to make many, often arbitrary methodological choices (e.g. parcellation scheme(s), connectivity model(s), tractography step size(s), etc.) that can greatly influence a network's configuration downstream. Because of the added model uncertainty that results, these untuned hyperparameters amount to a combinatorial explosion of 'hidden' researcher degrees of freedom that can easily distort statistical inference. Although perhaps previously thought to be an intractable problem, PyNets aims to directly address this methodological gap by offering a powerful computational framework for modeling individual structural and/or functional connectomes iteratively and with hyperparameter optimization.

Methods:

Despite its central concern with offering the user control over connectome-generating choices, PyNets strive for maximal user ease through end-to-end automaticity and easy installation (See: https://github.com/dPys/PyNets). To produce connectome estimates (i.e. as adjacency matrices or any of various embeddings) as output, it expects preprocessed neuroimaging files as input. Minimally, these file paths should include a preprocessed fMRI and/or dMRI image (or sets of images in the case of batch processing across subjects). Additional paths to a T1/T2-weighted anatomical images and brain masks can also optionally be specified to improve the quality of connectome estimation. As a post-processing workflow, PyNets further incorporates support for a variety of BIDS derivatives, and a formal BIDS API is currently under development.
Regardless of the inputs specified, PyNets strives by design to accommodate for all possible combinations of user input-that is, within a predefined set of constraints on execution, it will either complete successfully or return errors/warnings when estimation cannot be completed due to data or parameter incompatibility. These constraints include various hard-coded methods based on empirically-established standards such as the use of ensemble tractography and direct-streamline normalization. Beyond these static methods, however, the available connectome-generating hyperparameters in PyNets are extensive and cover aspects of both node and edge generation. Additionally, multiple values for each hyperparameter may be specified (e.g. multiple connectivity models, a threshold window, or several node sizes). In these cases, the PyNets workflow will execute much like grid-search, that automatically parallelizes itself optimally based on available compute resources.

Results:

Following connectome estimation, PyNets ultimately extracts a dataframe of graph topology metrics for each unique connectome estimate. PyNets also provides a variety of options for connectome embedding, community detection, multigraph estimation, visualization, and quality-control. It offers debugging tools, runtime statistics, and logging, along with other global configuration parameters controlling image orientation, resolution, and template type for advanced users. To accommodate for its high demands on IO, it has even recently incorporated SQL database support.

Conclusions:

The PyNets workflow offers an entirely new framework for connectome analysis. It has been tested on diverse neuroimaging data including resting-state and task-based fMRI, as well as dMRI acquired with varying directions, shells, and across scanners. It is fully-automated, meticulously optimized for time and space efficiency, capable of several layers of parallelized execution, scalable to large datasets, extensively tested, and under development to incorporate support for additional neuroimaging modalities.

Modeling and Analysis Methods:

Classification and Predictive Modeling
Connectivity (eg. functional, effective, structural) 2

Neuroinformatics and Data Sharing:

Workflows 1
Informatics Other

Keywords:

Computing
FUNCTIONAL MRI
Informatics
Machine Learning
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
Workflows
Other - Connectomics

1|2Indicates the priority used for review

My abstract is being submitted as a Software Demonstration.

Yes

Please indicate below if your study was a "resting state" or "task-activation” study.

Other

Healthy subjects only or patients (note that patient studies may also involve healthy subjects):

Patients

Are you Internal Review Board (IRB) certified? Please note: Failure to have IRB, if applicable will lead to automatic rejection of abstract.

No

Was any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.

Not applicable

Was any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.

Not applicable

Please indicate which methods were used in your research:

Functional MRI
Diffusion MRI
Computational modeling
Other, Please specify  -   Workflow programming, high-powered computing

Which processing packages did you use for your study?

Other, Please list  -   PyNets

Provide references using author date format

FILL IN