<!-- README.md is generated from README.Rmd. Please edit that file -->

RAVEL <img src="sticker.png" width = "150" align="right" />

Intensity normalization methods for structural MRIs

Creator: Jean-Philippe Fortin, fortin946@gmail.com

Authors: Jean-Philippe Fortin, John Muschelli

License: GPL-2

Software status

| Resource: | Travis CI | | ----------- | ---------------------------------------------------------------------------------------------------------------------------------------- | | Platform: | OSX | | R CMD check | <a href="https://travis-ci.org/Jfortin1/RAVEL"><img src="https://travis-ci.org/Jfortin1/RAVEL.svg?branch=master" alt="Build status"></a> |

References

| Method | Citation | Paper Link | | ----------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------- | | RAVEL | Jean-Philippe Fortin, Elizabeth M Sweeney, John Muschelli, Ciprian M Crainiceanu, Russell T Shinohara, Alzheimer’s Disease Neuroimaging Initiative, et al. Removing inter-subject technical variability in magnetic resonance imaging studies. NeuroImage, 132:198–212, 2016. | Link | | WhiteStripe | Russell T Shinohara, Elizabeth M Sweeney, Jeff Goldsmith, Navid Shiee, Farrah J Mateen, Peter A Calabresi, Samson Jarso, Dzung L Pham, Daniel S Reich, Ciprian M Crainiceanu, Australian Imaging Biomarkers Lifestyle Flagship Study of Ageing, and Alzheimer’s Disease Neuroimaging Initiative. Statistical normalization techniques for magnetic resonance imaging. Neuroimage Clin, 6:9–19, 2014. | Link |

Table of content

• 1. Introduction
• 2. Image preprocessing
• 3. Intensity normalization and RAVEL correction

1. Introduction

RAVEL is an R package that combines the preprocessing and statistical analysis of magnetic resonance imaging (MRI) datasets within one framework. Users can start with raw images in the NIfTI format, and end up with a variety of statistical results associated with voxels and regions of interest (ROI) in the brain. RAVEL stands for Removal of Artificial Voxel Effect by Linear regression, the main preprocessing function of the package that allows an effective removal of between-scan unwanted variation. We have shown in a recent paper that RAVEL improves significantly population-wide statistical inference. RAVEL is now part of the Neuroconductor project.

Installation

The following dependencies need to be installed first:

install.packages("devtools")
devtools::install_github("muschellij2/neurobase")
devtools::install_github("muschellij2/WhiteStripe")

RAVEL can be installed using the following command:

devtools::install_github("jfortin1/RAVEL")

2. Image preprocessing

We present a pre-normalization preprocessing pipeline implemented in the R software, from raw images to images ready for intensity normalization and statistical analysis. Once the images are preprocessed, users can apply their favorite intensity normalization and the scan-effect correction tool RAVEL as presented in Section 1 above. We present a preprocessing pipeline that uses the R packages ANTsR and fslr. While we have chosen to use a specific template space (JHU-MNI-ss), a specific registration (non-linear diffeomorphic registration) and a specific tissue segmentation (FSL FAST), users can choose other algorithms prior to intensity normalization and in order for RAVEL to work. The only requirement is that the images are registered to the same template space.

2.1. Prelude

To preprocess the images, we use the packages fslr and ANTsR. The package fslr is available on CRAN, and requires FSL to be installed on your machine; see the FSL website for installation. For ANTsR, we recommend to install the latest stable version available at the ANTsR GitHub page. The version used for this vignette was ANTsR_0.3.2.tgz. For the template space, we use the JHU-MNI-ss atlas (see Section 1.2) included in the EveTemplate package, available on GitHub at https://github.com/Jfortin1/EveTemplate. For data examples, we use 4 T1-w scans from the package RAVELData available on GitHub at https://github.com/Jfortin1/RAVELData. Once the packages are properly installed, we are ready to start our preprocessing of T1-w images. We first load the packages into R:

library(fslr)
library(ANTsR)
library(RAVELData)
library(EveTemplate)
have.fsl() # Should be TRUE if fsl is correctly installed

and let’s specify the path for the different files that we will need:

# JHU-MNI-ss template:
library(EveTemplate)
template_path <- getEvePath("T1")
# JHU-MNI-ss template brain mask:
template_brain_mask_path <- getEvePath("Brain_Mask")
# Example of T1-w MPRAGE image
scan_path <- system.file(package="RAVELData", "extdata/scan1.nii.gz")

2.2. JHU-MNI-ss template (_EVE_ atlas)

2.3. Registration to template

Tp perform a non-linear registration to the JHU-MNI-ss template, one can use the diffeomorphism algorithm via the ANTsR package. Note that we perform the registration with the skulls on. Here is an example where we register the scan1 from the RAVELData package to the JHU-MNI-ss template:

library(ANTsRCore)
library(ANTsR)
template    <- antsImageRead(template_path, 3)
scan <- antsImageRead(scan_path,3)
outprefix <- gsub(".nii.gz","",scan_path) # Prefix for the output files
output <- antsRegistration(fixed = template, moving = scan, typeofTransform = "SyN",  outprefix = outprefix)
scan_reg   <- antsImageClone(output$warpedmovout) # Registered brain

The object scan_reg contains the scan registed to the template. Note that the object is in the ANTsR format. Since I prefer to work with the oro.nifti package, which is compatible with flsr, I convert the object to a nifti object using the function ants2oro as follows:

# devtools::install_github("muschellij2/extrantsr")
# or
# source("https://neuroconductor.org/sites/default/files/neurocLite.R")
# neuro_install("extrantsr")
library(extrantsr)
scan_reg <- extrantsr::ants2oro(scan_reg)

We can save the registered brain in the NIfTi format using the writeNIfTI command:

writeNIfTI(scan_reg, "scan_reg")

Since scan_reg is converted to a nifti object, we can use the function ortho2 from the fslr package to visualize the scan:

ortho2(scan_reg, crosshairs=FALSE, mfrow=c(1,3), add.orient=FALSE)

2.4. Intensity inhomogeneity correction

We perform intensity inhomogeneity correction on the registered scan using the N4 Correction from the ANTsR package:

scan_reg <- extrantsr::oro2ants(scan_reg) # Convert to ANTsR object
scan_reg_n4 <- n4BiasFieldCorrection(scan_reg)
scan_reg_n4 <- extrantsr::ants2oro(scan_reg_n4) # Conversion to nifti object for further processing

2.5. Skull stripping

template_brain_mask <- readNIfTI(template_brain_mask_path, reorient=FALSE)
scan_reg_n4_brain <- niftiarr(scan_reg_n4, scan_reg_n4*template_brain_mask)
ortho2(scan_reg_n4_brain, crosshairs=FALSE, mfrow=c(1,3), add.orient=FALSE)

2.6. Tissue Segmentation

There are different tissue segmentation algorithms available in R. My favorite is the FSL FAST segmentation via the fslr package. Let’s produce the tissue segmentation for the scan_reg_n4_brain scan above:

ortho2(scan_reg_n4_brain, crosshairs=FALSE, mfrow=c(1,3), add.orient=FALSE, ylim=c(0,400))

The last line of code produces via the ortho2 function from the fslr package the following visualization of the template:

<p align="center"> <img src="images/template.png" width="600"/> </p>

We perform a 3-class tissue segmentation on the T1-w image with the FAST segmentation algorithm:

scan_reg_n4_brain_seg <- fast(scan_reg_n4_brain, verbose=FALSE, opts="-t 1 -n 3") 
ortho2(scan_reg_n4_brain_seg, crosshairs=FALSE, mfrow=c(1,3), add.orient=FALSE)

<p align="center"> <img src="images/seg.png" width="600"/> </p>

The object scan_reg_n4_brain_seg is an image that contains the segmentation labels 0,1,2 and 3 referring to Background, CSF, GM and WM voxels respectively.

2.7. Creation of a tissue mask

Suppose we want to create a mask for CSF.

scan_reg_n4_brain_csf_mask <- scan_reg_n4_brain_seg
scan_reg_n4_brain_csf_mask[scan_reg_n4_brain_csf_mask!=1] <- 0
ortho2(scan_reg_n4_brain_csf_mask