Neurodesk: An accessible, flexible, and portable data analysis environment for reproducible neuroimaging

Abstract

Neuroimaging data analysis often requires purpose-built software, which can be challenging to install and may produce different results across computing environments. Beyond being a roadblock to neuroscientists, these issues of accessibility and portability can hamper the reproducibility of neuroimaging data analysis pipelines. Here, we introduce the Neurodesk platform, which harnesses software containers to support a comprehensive and growing suite of neuroimaging software (https://www.neurodesk.org/). Neurodesk includes a browser-accessible virtual desktop environment and a command line interface, mediating access to containerized neuroimaging software libraries on various computing platforms, including personal and high-performance computers, cloud computing and Jupyter Notebooks. This community-oriented, open-source platform enables a paradigm shift for neuroimaging data analysis, allowing for accessible, flexible, fully reproducible, and portable data analysis pipelines.

Figure 1.

The Neurodesk platform. (a) The Neurodesk platform is built by and for the scientific community, enabling anyone to contribute recipes for new software containers to the repository. (b) Recipes contributed by the community are automatically used to build software containers and stored in the Neurocontainers repository. (c) Each software container packages a tool together with all the required runtime dependencies. The packaged software can therefore run identically in any supported computing environment. (d) Neurodesk provides two layers of accessibility: 1. Neurodesktop is a browser-accessible virtual desktop environment, allowing users to interact with the containerized software. 2. Neurocommand is a command-line interface that allows users to run the same software containers programmatically. These interfaces allow users to reproduce the same analysis pipelines across various computing environments.

Figure 2.

Discrepancies in image registration and tissue segmentation. (a) Calculation of the Dice dissimilarity coefficients; for each image, the voxel-wise disagreement in image intensity (FLIRT) or label (FIRST) calculated on System A vs System B was expressed as a proportion of the total number of voxels for each participant. (b) Histograms of Dice dissimilarity coefficients for image intensity calculated with FSL-FLIRT on Neurodesk vs. Local Install. To calculate these Dice coefficients, “disagreement” meant a voxel had a different intensity after image registration on System A vs. System B. Thus, the Dice coefficient of 0 for every participant whose images were registered using Neurodesk, means that the image intensity of each participant was matched across systems at every voxel. (c) Histograms of Dice dissimilarity coefficients for subcortical structure labels calculated using FSL-FIRST on Neurodesk vs. Local Install. To calculate these Dice coefficients, “disagreement” meant a voxel had different labels (e.g., amygdala, hippocampus, etc.) after image segmentation on System A vs. System B. Note that these Dice coefficients are much smaller than for image registration. This is expected because there are 73 times more “classes” for the image registration task, which uses image intensity (Range: 0 – 1903) as a class, than the classification task, which has labels for 15 structures. However, while both Neurodesk and the local system show strong agreement across systems overall, these distributions are completely non-overlapping, with Neurodesk showing much greater reliability across systems.

Figure 3.

Inter-system differences in image intensity in subcortical structures and subsequent classification of these subcortical structures. (a,b) Absolute voxel-wise differences in image intensity within subcortical structures after image registration with FSL-FLIRT on each system (i.e. ∣ Intensity_{system A} – Intensity_{system B}∣), averaged across participants. Projections are shown for image registration performed (a) using locally installed software, and (b) using Neurodesk (for which there were no intersystem differences). (c,d) Inter-system disagreement in subcortical structure labels after image segmentation with FSL-FIRST, averaged across participants. Projections are shown for image segmentation performed (c) using locally installed software and (d) using Neurodesk. (e) Scatter plot showing the mean inter-system image intensity differences across all voxels within the classified subcortical structures vs. the number of voxels subsequently classified with different labels across systems. For analyses performed with locally installed software, participants with larger differences in image intensity typically also had more prolific disagreement in labels between systems (Pearson’s r = 0.608, p < 0.001). This trend could not be assessed for Neurodesk, as there were no differences in image intensity across systems.

Figure 4.

Cumulative difference in the numbers of system library calls between System A and System B for the analysis run using the (a) locally installed and (b) Neurodesk version of FSL FIRST. Note that calls to floorf() were excluded from the plot as they occurred earlier in time and the discrepancies for floorf() far outnumbered those for any other function from the locally installed tool.