BENCH Practical

In this practical, you will discover how to use bench to investigate changes in brain microstructure between two subject groups using diffusion MRI. For an introduction to the fundamental concepts of change models, please refer to the BENCH paper or the BENCH training tutorial. Briefly, BENCH estimates the likelihood that a change in every parameter of a biophysical model can account for observed differences in the diffusion data.

Here we will employ BENCH to compare brain microstructure in 4 young subjects against 4 older subjects. Our goal is to identify regions in the brain where microstructural parameters vary with ageing. We use the standard model of white matter microstructure as our biophysical model.

Contents:

To conduct inference using BENCH, you will need to follow these steps:

Training Change Models

Using simulations to train machine learning models that can predict change in diffusion data

Extracting Summary Measurements

Estimating rotationally invariant features from cleaned diffusion MRI data

GLM

Running a General Linear Model to estimate the baseline values and changes in the summary measurements

Inference

Estimating the probability and amount of change in each parameter

Training Change Models

The first step in using BENCH is to train machine learning models known as "change models". These models aim to predict how changes in parameters of a biophysical forward model will affect diffusion MRI data. To train these models, BENCH utilizes simulations to generate the necessary training data. Once these change models are trained, they can be applied to investigate whether a specific parameter change can account for observed differences between groups. It is important to highlight that running this initial step does not require any diffusion data and only requires knowing the acquisition protocol and the forward biophysical model. To train a change model, you can execute the following command:

bench diff-train --model standard_model --bvals dataset/001/Diffusion/bvals --summarytype dt --output standard_model_dt 

In this command:

• diff-train: Specifies the sub-command, indicating that the action is to train a change model.
• --model standard_model: Sets the microstructural forward model according to those available in the BENCH library. The choice of forward model depends on the tissue type and parameters of interest.
• --bvals dataset/001/Diffusion/bvals: Specifies the diffusion protocol using the b-values file from one subject. This is assumed to be consistent across all subjects.
• --summarytype dt: Indicates that the change model should be trained on diffusion tensor metrics (MD and FA per shell) as summary measurements.
• --output standard_model_dt: Defines the location where the weights for the trained models are saved.

For an in-depth explanation of these options or to learn how to implement your own specific forward model, please see the BENCH training tutorial.

Training a change model can be a time-consuming process, potentially taking several hours. The duration depends on factors such as available computational resources, the specifics of the diffusion protocol, the number of samples used for training, etc. However, once the model is trained, it can be reused on any dataset that employs the same diffusion acquisition protocol, saving time in future analyses. To save time, we provide a pre-trained model for this acquisition which you can use for this tutorial.

Extracting summary measurements

Using diffusion MRI signals directly to investigate brain microstructural changes presents two challenges: the high dimensionality of the data and the complexity arising from tissue orientation and acquisition protocol. BENCH addresses these issues by employing signal "representations", e.g., Diffusion Tensor or Spherical Harmonics. These models transform the diffusion data into summary measurements to provide metrics that are invariant to tissue orientation, thus simplifying analysis.

For example, metrics such as mean diffusivity and fractional anisotropy can be used to summarize the diffusion data within each shell. These metrics offer the advantage of being rotationally invariant, meaning they remain consistent regardless of the tissue orientation or the orientation of head in the acquisition.

To extract summary measurements from the diffusion MRI data, use the following command after changing the directory to the bench folder:

cd ~/fsl_course_data/bench

bench diff-summary --data dataset/001/Diffusion/data.nii.gz \
                   --bvecs dataset/001/Diffusion/bvecs \
                   --bvals dataset/001/Diffusion/bvals \
                   --mask wm_mask_2mm.nii.gz \
                   --xfm dataset/001/xfms/diff2standard_warp.nii.gz \
                   --summarytype dt \
                   --output summaries/001

In this command:

• diff-summary: Specifies the sub-command for extracting summary measurements from diffusion MRI data.
• --data, --bvals, --bvecs: Identify the diffusion MRI dataset and its associated acquisition protocol. The data must have been pre-processed (e.g., using eddy/topup)
• --xfm: Provides the transformation for mapping the diffusion data to a standard space. Here we use MNI space.
• --mask wm_mask_2mm.nii.gz: Specifies which regions of the brain to target for summary measurement extraction. It should be provided in the standard space. In our example, we focus on white matter by providing a white matter mask.
• --summarytype dt: Indicates the types of summary measurements to extract. For this tutorial, we use the diffusion tensor representation, i.e. mean diffusivity and fractional anisotropy separately for each shell.
• --output: Defines where the summary measurements are saved.

It is worth mentioning that BENCH uses nearest-neighbor sampling to compute summary measurements for each voxel in the mask from the diffusion data. This approach avoids the need for spatial transformations, which can be costly for diffusion data, and also eliminates the need to adjust b-vectors.

Now open one of the subject's summary measurement files in FSLEYES using the following command:

fsleyes summaries/001/* &

Summary Extraction for Multiple Subjects

For single-subject processing, you would execute the above command. However, to compute summary measurements for all 8 subjects in our example, you can use a bash for loop:

for i in $(seq 1 8); do
    bench diff-summary --data dataset/00$i/Diffusion/data.nii.gz \
                       --bvecs dataset/00$i/Diffusion/bvecs \
                       --bvals dataset/00$i/Diffusion/bvals \
                       --mask wm_mask_2mm.nii.gz \
                       --xfm dataset/00$i/xfms/diff2standard_warp.nii.gz \
                       --summarytype dt \
                       --output summaries/00$i
done

Running the GLM

After extracting these rotationally invariant summary measurements for each subject, ensuring they are all in the template space, you can proceed to run a General Linear Model (GLM) analysis. This will allow you to calculate the baseline signal, the differences between the two groups, and the covariance of the noise. Armed with this information, you will be ready to quantify how the microstructural changes relate to changes in the diffusion data between the groups.

To perform a GLM analysis, use the following command:

bench glm --summarydir summaries \
          --mask wm_mask_2mm.nii.gz \
          --designmat design.mat \
          --designcon design.con \
          --subjectlist subjlist.txt \
          --output Glm

In this command:

• glm: Specifies that a GLM analysis will be run.
• --summarydir summaries: Points to the directory containing summary measurements for all subjects, organized into subdirectories per subject that contain all summary measurement images.
• --mask wm_mask_2mm.nii.gz: Defines the regions of interest for the GLM.
• --designmat design.mat: Provide the design matrix in FSL format. The design matrix should have one row per subject and can have multiple columns.
• --designcon design.con: In the design contrast the first contrast should represent the group mean and the second contrast should indicate the group difference.
• --subjectlist subjlist.txt: Lists the subjects involved in the analysis one per line, the names must match the subdirectories in the summary dir and the order must match the design matrix.
• --output Glm: Specifies where the GLM results will be saved.

Now you can explore the changes in summary measurements across the brain attributed to group differences using the following command:

fsleyes Glm/change.nii.gz --cmap red-yellow --negativeCmap blue-lightblue --useNegativeCmap -dr 0 .1 &

Change the volume number and modify the data range as needed to examine how each summary measurement has changed. The corresponding names of the summary measurements can be found in the text file ./Glm/summarynames.txt.

Running Inference

Now we are set for the main analysis. We have estimates of how the diffusion data varies between our two groups, namely old versus young. Additionally, we have trained machine-learning models using simulations to understand how changes in microstructural parameters can influence the data. By combining these elements, we can explore which types of changes in our data across the brain best align with which of these trained change models.

To execute the inference step, use the following command:

bench inference --changemodel standard_model_dt --glmdir Glm --mask slice_mask_2mm.nii.gz --output Results

In this command:

• inference: Specifies that inference will be conducted.
• --model standard_model_dt: Points to the file containing the change models that were trained in the first stage.
• --glmdir Glm: Provides the directory containing GLM results, including group means, group differences, and noise covariance for all summary measurements.
• --mask slice_mask.nii.gz: Provides the brain mask to specify voxels that are included in the inference step. Here we use a slice mask for time considerations.
• --output Results: The location where the resulting probability maps will be saved.

This step generates two key outputs for each parameter of the forward model:

1. A probability map that displays, for each voxel, the normalised likelihood that the parameter can explain the observed differences in the diffusion MRI data between the groups.
2. An estimate of the most likely amount of change for that parameter.

These outputs help to quantify and localize where specific microstructural changes can account for observed group differences.

Now you can look at the final probability maps to see to what extent each parameter can explain the group differences in the summary measurements from diffusion MRI data using the following command:

fsleyes Results/*_probability.nii.gz --cmap red-yellow -dr 0 1 &

It is also possible to check the most likely amount of change in each parameter using the following command:

fsleyes Results/*_amount.nii.gz --cmap red-yellow --negativeCmap blue-lightblue --useNegativeCmap &

Moreover, in order to see what parameter change can best explain differences between the groups you can run the following command:

fsleyes Results/inferred_change.nii.gz --cmap random &

The End.