Reproducibility of neuroimaging analyses across operating systems

Abstract

Neuroimaging pipelines are known to generate different results depending on the computing platform where they are compiled and executed. We quantify these differences for brain tissue classification, fMRI analysis, and cortical thickness (CT) extraction, using three of the main neuroimaging packages (FSL, Freesurfer and CIVET) and different versions of GNU/Linux. We also identify some causes of these differences using library and system call interception. We find that these packages use mathematical functions based on single-precision floating-point arithmetic whose implementations in operating systems continue to evolve. While these differences have little or no impact on simple analysis pipelines such as brain extraction and cortical tissue classification, their accumulation creates important differences in longer pipelines such as subcortical tissue classification, fMRI analysis, and cortical thickness extraction. With FSL, most Dice coefficients between subcortical classifications obtained on different operating systems remain above 0.9, but values as low as 0.59 are observed. Independent component analyses (ICA) of fMRI data differ between operating systems in one third of the tested subjects, due to differences in motion correction. With Freesurfer and CIVET, in some brain regions we find an effect of build or operating system on cortical thickness. A first step to correct these reproducibility issues would be to use more precise representations of floating-point numbers in the critical sections of the pipelines. The numerical stability of pipelines should also be reviewed.

Keywords: CIVET; FSL; Freesurfer; operating systems; reproducibility.

Figure 1

Source code, compilation, libraries, kernel and hardware.

Figure 2

Sum of binarized differences between cortical tissue classifications obtained on cluster A and cluster B (FSL FAST, build 1, n = 150 subjects). All binarized differences were resampled to the default MNI152 volume template.

Figure 3

Cumulative inter-OS differences in FSL FAST.

Figure 4

Histograms of Dice coefficients between classifications obtained on cluster A vs. cluster B with FSL FIRST. All bins are of size 0.1. min, μ, and σ are the minimum, mean and standard deviation Dice coefficient, respectively.

Figure 5

Sample subcortical classifications with FSL FIRST: subject 260, Z = 114.

Figure 6

Cumulative inter-OS differences in FSL FIRST.

Figure 7

Histograms of Dice coefficients between matched ICA components. Variable random seed initialization (case b), automatic dimension detection (case d). Red histograms, positive components; Blue histograms, negative components; m, mode of the histogram.

Figure 8

Histograms of Dice coefficients between matched ICA components. Variable random seed initialization (case b), fixed dimension (case c). Red histograms, positive components; Blue histograms, negative components; m, mode of the histogram.

Figure 9

Histograms of Dice coefficients between matched ICA components on cluster A vs. cluster B (logarithmic scale). Fixed random seed initialization (case a). Red histograms, positive components; Blue histograms, negative components; m, mode of the histogram.

Figure 10

Mean absolute differences after successive steps of pre-processing, normalized by the mean absolute difference after all pre-processing steps (all 37 subjects).

Figure 11

Cumulative inter-OS differences in FSL mcflirt.

Figure 12

Surface maps of mean absolute difference, standard-deviation of absolute difference, t-statistics and RFT significance values showing regions where the cortical thickness extracted with Freesurfer differs for build 1 and build 2 (both executed on cluster B).

Figure 13

Surface maps of mean absolute difference, standard-deviation of absolute difference, t-statistics and RFT significance values showing regions where the cortical thickness extracted with Freesurfer differs for cluster A and cluster B (both executed with build 1).

Figure 14

Surface maps of mean absolute difference, standard-deviation of absolute difference, t-statistics and RFT significance values showing regions where the cortical thickness extracted with CIVET differs for cluster A and cluster B (both executed with build 1).