DTI:Acquisition

Definition of DTI Protocols

Temporary page containing content from the Morphometry BIRN Best Practices DTI page that is being updated.
Back to SFN2009_Slicer_DTI_Workshop

Work is ongoing within the BIRN project to define DTI acquisition protocols that can be used in single and multi-site trials. This work is being informed by the DTI calibration studies performed by the JHU group.

Basic Steps for Processing DTI data

1) If required, convert data from the original format to another file format that your software accepts. Some common ones are:

• Digital Imaging and Communications in Medicine (DICOM) - all scanners that I know of support DICOM, but one must be aware that there are variations in the way that DWI information is stored and important information may be stored in the vendors private fields. The relevant fields for DWI are:
  • 0018 9075 CS 1 Diffusion Directionality
  • 0018 9076 SQ 1 Diffusion Gradient Direction Sequence
  • 0018 9087 FD 1 Diffusion b-value
  • 0018 9089 FD 3 Diffusion Gradient Orientation
  • 0018 9117 SQ 1 MR Diffusion Sequence
  • 0018 9147 CS 1 Diffusion Anisotropy Type

Note that while there are tags for the gradient orientation and b-value there is not a tag for the entire b matrix. Also notice that the vendor-specific measurement frame is also not stored - essentially you need to know the relationship between the coordinate frame in which the gradient directions are defined and the coordinate frame in which the image orientation is defined or your color maps will not be correct. Since, one usually doesn't know these coordinate frames, one must resort to flipping the sign of various gradient direction components until the directions given by the tensor processing correspond to what you know from the anatomy. The NA-MIC group has also noted that GE has stored the gradient directions, not in the tag assigned to that quantity, but in other fields. Siemens (as of 2006) stores the b-matrix as well as other diffusion information.

See here for more information about the format of DTI data as it is acquired by each of the dominant MR scan manufacturers. Learn more about DICOM here.

• Nearly Raw Raster Data (NRRD) - see teem.sourceforge.net/nrrd/format.html
• Neuroimaging Informatics Technology Initiative (NIfTI, nifti.nimh.nih.gov/) - this format has a 352 byte header
• Raw (no header) - DTIStudio allows the user to input all the file parameters using a .dpf (diffusion parameters file) file
• Analyze - This format uses either separate files for the header and images or an image file with a header. The extension is .hdr / .img The .img file contains the images. ( www.grahamwideman.com/gw/brain/analyze/formatdoc.htm )
• PAR/REC - this is a Philips format. The .PAR file contains the parameters for the scans and the .REC file contains the data.
• p-file - this is a GE format, which contains the images of interest (and possibly other files). The data is stored in the /usr/g/mrraw directory on the scanner and have the format PXXXXXX.X

2) If necessary for your processing application, make files that contain the b-values and b-vectors for your data. Determine whether the b=0 scans come at the beginning or the end of the data file. Are there any extra frames included in the volume, i.e., some manufacturers include an 'isotropic-weighted' scan at the end of the volume (Philips does this, see below)? If so, you should remove these files before you convert the original files to the format you will be working with.

3) Perform self-registration of the data set. With structural/T1-weighted data you could do rigid alignment of the data(6 degree-of-freedom {DOF} registration), but with DTI data it is better to do a 12 DOF registration since each unique diffusion-weighted direction (DWD) results in different image distortions. There are existing tools available to correction for these distortions, for example, the tool here is a BIRN developed tool.

4) Check the image volume for bad datasets. "Bad" data usually arises from subject motion. Given that these scans are diffusion sensitized, subject motion can affect the signal intensity in an image. If the motion is large enough image intensity dropout can occur. Some tensor calculation packages allow you to remove from the calculation data that is thought to be bad (DTIStudio allows you to do this, for example). If your tensor calculation package does not have this feature, lobby the code developers to include it!

5) Calculate the tensors and associated metrics (FA, eignevalues, eigenvectors, etc.) using your processing package.

6) Use Brain Extraction Tool (BET, University of Oxford FMRIB Software Library, www.fmrib.ox.ac.uk/analysis/research/bet/ ) or some other method to get rid of the noise pixels around the brain. Some packages use a simple erosion-based algorithm. Personally, I use BET to get rid of most of the unwanted pixels and then my own code to clean up the bit that BET leaves behind. This is most important if you are doing something like whole-brain histograms of FA, but not important at all if you are doing ROI analysis.

7) Analyze the tensor metric data. At this point you can construct color maps from the 1st eigenvalue and FA maps, perform fiber tracking, etc.

Vendor Specific Information

Philips (last updated 3/1/09)

• Philips data can come in DICOM or PAR/REC format. The PAR is the paramter file which gives a complete listing of the paramters used for each scan. The REC file is a volume that contains the images. This file has no header. A matlab-based package called CATNAP (Coregistration, Adjustment, and Tensor-solving, a Nicely Automated Program, can be used to convert the PAR/REC files to Analyze format (among other nice features it has).
  
• Philips gives you an extra scan - an isotropic-weighted scan at the end of the REC file. This scan is not used in the DTI processing so you have to rename or remove it before you convert the original files to the format you will be working with.
• The current version puts the b=0 scans at the end of the volume. Older versions put them at the beginning of the volume.
  
  

Siemens (last updated 3/1/09)

• To use a non-Siemens set of diffusion-weighted directions on a scanner running vb15:
  

You will have to switch a file on the scanner to use a custom file with the user-defined gradient directions in it. In the "Free" mode, the ASCII file on the scanner defines the DTI directions and is read by the sequence itself. The text file is located in C:MedComMriCustomerseq. Rename the current DiffusionVectors.txt file to DiffusionVectors_orig.txt first and then rename DiffusionVectors_user.txt to DiffusionVectors.txt (Obviously you then have to put back the original file after your run.) Then, to get the scanner to recognize that you want to use a file other than the default, go into the Diff tab when you are editing the DTI sequence parameters and change the default "MDDW" to "Free".

GE (last updated 3/4/09)

• GE currently has two dummy scans as the default, so the first scan may have an increased signal intensity relative to the others. This can be mitigated by increasing the repetition time (TR), but (assuming that the first set of scans are the b=0 scans) it is better to take additional b=0 scans and throw away the first few. This ensures that the intensity has reached a steady state for the b=0 volumes that will be used.
• If partial Fourier imaging is enabled, the default number of lines acquired past k=(0,0) is 6. This may not be enough for the heterogeneous tissue at the bottom of the brain and can result in an alternating image intensity (i.e., bright/dark) for the lowermost slices in the volume. This artifact makes the data very difficult to self register and gives bad eigenvalues, etc. Because of this, we recommend not to use partial Fourier imaging.
  
  

Toshiba (last updated 3/1/09)

• I've only processed a single data set from this manufacturer. However, it seems as though they intersperse the b=0 scans throughout the data volume. The set I processed was
  

2 b=0
10 b NE 0
2 b=0
10 b NE 0
etc…
I am not sure whether this is standard or the site just set up the scan order that way.

General Comments for Multi-site trials

In the case of diffusion tensor imaging, "optimization" of the protocol depends upon the brain structure of interest. In general, the lower the fractional anisotropy (FA) in a given region, the more scan-time units will be needed to accurately calculate the FA. The Johns Hopkins University MBIRN site has performed DTI calibration studies to identify salient issues in DTI data acquisition. They discuss the amount of DTI data needed to resolved certain brain structures in terms of the "scan-time unit", which is defined as a single run consisting of 5 b=0 scans, and 30 diffusion-weighted directions with a single average. This work is discussed in the Farrell paper cited below.

That study found that 3 scan-time units are sufficient for good accuracy and reliability of DTI measurements for a usual set of brain structures. That means that you can either repeat a 5 b=0, 30-direction DTI scan 3 times or repeat a 3 b=0, 15-direction DTI scan 6 times. Note that in the latter example the number of b=0 images is close, but not identical to the former case since you cannot specify half of a scan. However, note that repeating scans with a smaller number of directions may bias your diffusion measurement. The conditions under which this can occur is discussed in the Landman paper cited below.

General Comments

• Use a standard set of diffusion-weighted directions, like Jones 30 for example, since the DWD's may be different between different manufacturers. For example, the default DWD's for Siemens and GE are not the same for a given number of DWD's. Most vendors allow you to enter your own set of DWD's or, as is the case with Siemens, you can choose a different scan mode that will then read in a file with the non-vendor DWD's. You need to check however, on your DTI license. As the scanner software version change, the number of DWD's you can specify with the standard license may also change. In some cases, you may need a separate or research license to use a large number of DWD's.
• A good rule of thumb is that at 1.5T you need 2.5 mm^3 voxels to get good DTI results. At 3.0T you can reduce the voxel size to 2.0 mm^3.
• The brain structure with the lowest FA that you are interested in determines how many b=0 scans and diffusion-weighted directions you need to acquire. See the paper by Farrell et al., 2007 below. This paper shows curves how many scan-time units (1 STU = 5 b=0 + 30 b NE 0 scans) are needed to achieve accurate and reproducible FA values.

A reasonable protocol to use to determine the quality of DTI data at multiple sites is:
Field Strength: 1.5 T
Image matrix (raw data points): 96 x 96
Reconstructed matrix: 96 x 96 (no interpolation is better, but interp to 192x192 ok)
K-space coverage: full/symmetric - don't do partial fourier k-space coverage
Field of view: 240 x 240 mm
Slice thickness: 2.5 mm (no gap)
The number of slice: 25
Parallel imaging: SENSE (p = 2) for Philips and GRAPPA for SIEMENS
TE: Shortest (Be aware that the shortest TE sometimes causes lengthening of TR due to a duty cycle problem and significantly increases scanning time. Adding 4 - 8 ms to the minimum TE should solve this problem)
TR: Shortest to accommodate all 25 slices. Typically ~4 sec
Cardiac gating: No
Signal averaging: No signal averaging
b-value: 1000 s/mm^2
Gradient orientation scheme: Jones 30 + 5 b0 (the is one scan-time-unit STU)

This should take about 2:30 min for each scan. The idea here is to look at how FA in various structures changes with the number of STU. Acquire 6-10 runs of the above protocol. The data is then concatenated before tensor processing in the following manner to form data sets with increasing numbers of STU (and hence signal-to-noise ratio)
set 1
sets 1+2
sets 1+2+3
sets 1+2+3+4
etc.
One can then use these curves to determine how the accuracy and reproducibility of the FA of various brain structures changes with STU and compare the curves for each site to look for outliers. This method is detailed in Farrell et al., 2007 (below).

General References for DTI

1. J. A. Farrell, B. A. Landman, C. K. Jones, S. A. Smith, J. L. Prince, P. C. van Zijl, and S. Mori. "Effects of SNR on the Accuracy and Reproducibility of DTI-derived Fractional Anisotropy, Mean Diffusivity, and Principal Eigenvector Measurements at 1.5T", Journal of Magnetic Resonance Imaging. 26(3): 756-767. August 2007.
2. B. A. Landman, J. A. Farrell, C. K. Jones, S. A. Smith, J. L. Prince, P. C. van Zijl, and S. Mori. "Effects of Diffusion Weighting Schemes on the Reproducibility of DTI-derived Fractional Anisotropy, Mean Diffusivity, and Principal Eigenvector Measurements at 1.5T", NeuroImage. 36(4): 1123-1138. July 2007.
   

Back to SFN2009_Slicer_DTI_Workshop