This repository contains code and data used in the following publication:
J. Gomez-Ramirez et al, "A comparative analysis of automated MRI brain segmentation in a large longitudinal dataset of elderly subjects" (pre-print on BioRxiv: https://doi.org/10.1101/2020.08.13.249474 )
Abstract
In this study, we perform a comparative analysis of automated image segmentation of subcortical structures in the elderly brain. Manual segmentation is very time-consuming and automated methods are thus, gaining importance as a clinical tool for diagnosis. The two most commonly used software libraries for brain segmentation -FreeSurfer and FSL- are put to work in a large dataset of 4,028 magnetic resonance imaging (MRI) scans collected for this study. We find a lack of linear correlation between the segmentation volume estimates obtained from FreeSurfer and FSL. On the other hand, FreeSurfer volume estimates tend to be larger than FSL estimates of the areas putamen, thalamus, amygdala, caudate, pallidum, hippocampus, and accumbens. The characterization of the performance of brain segmentation algorithms in large datasets as the one presented here, is a necessary step towards partially or fully automated end-to-end neuroimaging workflow both in clinical and research settings.
Dataset description
The dataset contains two csv files:
- df_fsl_lon.csv is the Pandas dataframe containing the results of the automated segmentation performed with FSL
- df_free_lon.csv contains the Pandas dataframe containing the results of the automated segmentation performed with FreeSurfer.
The fields include in the dataset are as follows:
- Age the age of the participant in the moemnt of performing the MRI scan (%.2f)
- Sex encoded as 0 Male and 1 Female
- Subcortical Volume estimates use the nomenclature: [fsl|free] [R|L] [structure] where structure can be Thalamus, Accumbens, Pallidum, Hippocampus, Amygdala, Caudate and Putamen. The volume is expressed in mm^3.
df_fs_lon.csv.shape
[7080 rows x 16 columns]
df_fsl_lon.columns
Index(['age', 'sex', 'fsl_R_Thal', 'fsl_L_Thal', 'fsl_R_Puta', 'fsl_L_Puta',
'fsl_R_Amyg', 'fsl_L_Amyg', 'fsl_R_Pall', 'fsl_L_Pall', 'fsl_R_Caud',
'fsl_L_Caud', 'fsl_R_Hipp', 'fsl_L_Hipp', 'fsl_R_Accu', 'fsl_L_Accu'],
dtype='object')
df_free_lon.shape
[7080 rows x 16 columns]
df_free_lon.columns
Index(['age', 'sex', 'free_R_Thal', 'free_L_Thal', 'free_R_Puta',
'free_L_Puta', 'free_R_Amyg', 'free_L_Amyg', 'free_R_Pall',
'free_L_Pall', 'free_R_Caud', 'free_L_Caud', 'free_R_Hipp',
'free_L_Hipp', 'free_R_Accu', 'free_L_Accu'],
dtype='object')
MRI Data collection A total of 4028 MRIs were collected in 5 years, 990 in the first visit, 768 in the second, 723 in the third, 634 in the fourth, 542 in the fifth, and 371 in the sixth year. The imaging data were acquired on a 3T General Electric scanner (GE Milwaukee) utilizing the following T1-weighted inversion recovery, flip angle 12°, 3-D pulse sequence: echo time Min. full, time inversion 600 ms, Receiver Bandwidth = 19.23 kHz, field of view = 24.0 cm, slice thickness = 1 mm, Freq. x Phase = 288 x 288. The preprocessing of MRI 3 Tesla images in this study consisted of generating an isotropic brain image with non-brain tissue removed. We used the initial, preprocessing step in the two computational segmentation tool used in this study: FSL pipeline (fsl-anat) and the FreeSurfer pipeline (recon-all). We run both pipelines in an identical computational setting: Operating System Mac OS X, product version 10.14.5 and build version 18F132. The version of FreeSurfer is FreeSurfer-darwin-OSX-ElCapitan-dev-20190328-6241d26. The version of the BET tool for FSL is v2.1 - FMRIB Analysis Group, Oxford and the FIRST tool version is 6.0.
[FreeSurfer, 2017] FreeSurfer cortical reconstruction and parcellation process. (2017).Anatomical processing script:recon-all. https://surfer.nmr.mgh.harvard.edu/fswiki/recon-all.
[FSL, 2017] FSL (2017). Anatomical processing script: fsl_anat. https://fsl.fmrib.ox.ac.uk/ fsl/fslwiki/fsl_anat.