Python-based Reliability in MRI (PyReliMRI)
PyReliMRI is described and applied in the TBD Preprint.
Intro of Problem
Reliability questions for task fMRI and resting state fMRI are increasing. As described in 2010, there are various ways that researchers calculate reliability. Few open-source packages exist to calculate multiple individual and group reliability metrics within one place one.
Purpose of Package
The purpose of the PyReliMRI is to provide an open-source python package that will estimate multiple reliability metrics on fMRI (or MRI) data in standard space – at the group and individual level – that researchers may use to report in their manuscripts in cases of multi-run and/or multi-session MRI data. If you don’t have multi-session or multi-run data, some of the features of this packages can still be useful!
Example 1: you may have single session and single run task-fMRI data and you can split the run and model them separately. In this session, you can calculate reliability or similarity metrics on these files.
Example 2: you may have group level maps (e.g., within or across studies for specific contrasts) or have access to neurovault data. You can use these to calculate various similarity metrics.
PyReliMRI is composed of a series of modules, each of which correspond to different use cases. So there are numerous questions that can be answered on the same data. The package is intended to be used with 3D brain images that are in standard space.For example, Nifti fMRI brain maps in MNI or Talairach space. However, in other fields it may be defined differently.At the group level (this doesnt have to be solely used for group level maps, but the common option), the functions in the similarity module calculate the similarity between two 3D Nifti images using Dice or Jaccard similarity coefficients, or tetrachoric or spearman correlation. In addition to calculating the similarity between two NifTi images a pairwise_similarity option is available to calculate pairwise similarity coefficients across a list of 3D Nifti images and return a list of coefficients with associated image labels. The latter option is in efficient way to extract similarity metrics across your list of images in a single shot.
At the individual level, the functions in the brain_icc module calculate intraclass correlations. For description of different ICCs and their calculations, see discussion in Liljequist et al., 2019 (for conceptual issues in fMRI, see Noble et al., 2021 . In this package, you have the option to select ICC(1), ICC(2,1) or ICC(3,1). The brain_icc module contains an option to calculate voxelwise ICC and ROI based ICCs. The ROI based ICC is integrated with the Nilearn datasets. As a result, the atlas options include deterministic: AAL, Destrieux 2009, Harvard-Oxford, Juelich, Pauli 2017, Shaefer 2018, Talairach, and probablistic options: Difumo, Harvard-Oxford, Juelich, Pauli 2017 and Smith 2009. Take note of the quality of each atlas as it is uploaded to Nilearn Datasets and confirm it aligns with your project goals. Some coverage across MNI brain maps may vary (e.g., Juelich and Talairach) and probabilistic atlases will have a certain level of smoothing as a result of masking the data by the atlas.
Citation
If you use PyReliMRI in your research, please cite the following Zenodo DOI:
Demidenko, M. I., & Poldrack, R. A. (2023). PyReliMRI: An Open-source Python tool for Estimates of Reliability in MRI Data [Computer software]. https://zenodo.org/record/8387971