Diffusion MRI acqusition


Currently there is no consensus on the "optimal" b-table setting to acquire diffusion MRI for "beyond-DTI" analysis. 10 years ago, the trend in the field was "high angular resolution diffusion imaging" (HARDI), which acquires hundreds of sampling at one b-value (typically 3000 or more), but then HARDI was gradually replaced by multishell acquisition, which acquired diffusion MRI using at least two different b-values. This trend was motivated by studies confirming the benefit of combining data from multiple b-values. 

Nonetheless, multishell acquisition has optimization problems. For example, the human connectome project acquired 90-90-90 directions at b-values of 1000, 2000, and 3000.  These 90 sampling directions are not equally distributed within each shell. Some directions have a higher sampling density, whereas others have lower. Consequently, the scheme has a lot of "rotation variation". Moreover, the 90 directions at lower b-value could be over-sampling, whereas the 90 directions at high b-value could be under-sampling. This is evident if we calculate the correlation between any two DWI in each shell, the correlation is much higher in the low b shell, suggesting greater data redundancy.

10-minute q-space acquisition

My personal recommendation is a 10-minute "grid-258" sampling with maximum b-value of 4000, which acquires, not just two or three b-values, but 23 different b-values ranging from b=0 to b=4000 at a total of 258 directions. Using a multi-band sequence (e.g. CMRR) with an MB factor of 4, this 2-mm dMRI acquisition can be done in around 10 minutes. A 5-minute version is also available using "grid-101".

The grid sampling has several benefits:
  1. It has uniformly distributed density in the diffusion encoding space (i.e. q-space). This avoids over-sampling at low b-values or under-sampling at high b-values. It does not have the sampling homogeneity problem in the shell acquisitions.
  2. It can be reconstructed by DTI, ball-and-sticks model, NODDI, GQI...etc.
  3. It captures a continuous range of diffusion patterns from none-restricted diffusion to restricted diffusion. For clinical studies, the grid scheme can capture all possible diffusion change due to edematous tissue or cell infiltration. in comparison, multi-shell only acquires 2 or 3 b-values and may miss diffusion patterns that are only sensitive to values in between.
There are limitations with the grid sampling scheme:
  1. It's Eddy current artifact cannot be corrected using FSL eddy, and the bipolar pulse is often needed to handle eddy current at the sequence level
  2. Methods using spherical harmonics cannot use grid schemes because there is no shell structure for estimating spherical harmonics.
The following steps will help you set up the grid-258 sampling scheme on your MRI scanner. The following steps are verified on Siemens scanners, and a similar protocol can be implemented in other manufacturers.

STEP1: Download the b-table

If you are using a Siemens Scanner, you may download the vector table here: 

101 directions on grid: https://pitt.box.com/v/GRID101-BTABLE
128 directions on grid: https://pitt.box.com/v/GRID128-BTABLE
258 directions on grid: https://pitt.box.com/v/GRID258-BTABLE

If you are using other Scanners, you may need to convert the b-table to its compatible format: https://pitt.box.com/v/GRID258

STEP2: Install a multi-band EPI sequence

The MRI scanner needs to have a multi-band diffusion sequence installed. For example, the CMRR multi-band sequence: http://www.cmrr.umn.edu/multiband/ or you may use any multi-band sequence provided by the MRI vendor. There could be a licensing fee with the multi-band sequence.

*Make sure that the sequence allows for loading an external b-table (e.g. the "Free" mode)

STEP3: Setup parameters

Set up the parameters for the sequence in the following steps:

1. In-plane resolution: 2.0 mm, slice thickness: 2.0 mm (if your SNR is not good enough, increase them to 2.4 mm.)
2. Matrix size: 104x104
3. Slice number: 72 (can be reduced if ignoring the cerebellum), no gap
4. Multi-band acceleration factor: 3 or 4
5. In the diffusion tab, load the vector table obtained from STEP1 by "Free" mode.
6. b-value1=0 and b-value2=4000
7. "Bipolar" diffusion scheme (for eliminating eddy current). Some may prefer "Monopolar" and correct Eddy current using FSL's eddy. This only works on shell acquisition, and I do not recommend this approach for grid scheme.
8. Minimum TE and TR
9. Pixel bandwidth: ~1700
10. Phase encoding direction: A to P
11. Make a copy of the sequence, invert its phase encoding direction (P to A). Only b0 is needed here for phase distortion correction. 

STEP4: Quality Check

1. Make sure that you can still see the brain contour in the DWI with b=4000. If not, consider lowering the b-value to 3000.
2. Create SRC files from the diffusion MRI data. Run DSI Studio and use [Tool: Batch Processing][SRC Quality control] to select a folder that contains the SRC file. It will compute "Neighboring DWI correlation". The one with a low correlation value may indicate a problem in data acquisition.

*Please feel free to send me your grid258 data for a quality check. I will compare the results with the data I have to make sure that you have achieved the same quality.

What if I only have the default DTI protocol? 

The good news is that you can use scanner's built-in DTI protocol to acquire "multishell" and still enjoy the benefit of "beyond-DTI" methods such as GQI, QSDR, RDI...etc.

Here's a working setting on a SIEMENS 3T Scanner:

1. acquire one 32-direction DTI at b=1500 and another 60-direction DTI at b=3000 (built-in ep2d_mddw protocol)
2. In-plane resolution: 2mm, slice thickness: 2mm
3. Matrix size: 104x104
4. Slice number: 72 (can be reduced if ignoring the cerebellum), no gap
5. Minimum TE and TR, but the TE for the b=1500 DTI should be the same as the TE of the b=3000 DTI.
6. Make a copy of the sequence and invert its phase encoding direction (acquire AP and PA for phase distortion correction)

In the analysis, copy the DICOM files from these two DTI acquisitions together.

This acquisition will require FSL's eddy to correct for the eddy correct artifact.

An important issue about motion correction

FSL's eddy is the best solution for motion correction, and I highly recommend using it if it accepts your data.

However, it is worth noting that the correction is using "data redundancy" to replace corrupted data. To be specific, motion correction routine first discards corrupted DWI volume/slices, and then the missing slices were estimated using "nearby DWI" (DWI with a similar diffusion gradient encoding). Motion correction works means that (1) several DWI samples are redundant (2) you can acquire less DWI samples in a shorter scanning time to get a similar quality. 

An optimized sampling will have redundancy minimized, and there is no chance for motion correction. The grid sampling scheme I suggested above is close to this optimized condition. FSL's eddy will not improve its quality because there is minimal data redundancy. 

My past experience is that if there is visible motion in the acquisition, the only choice is to discard the entire scan and redo it. DSI Studio has a routine for DWI quality check and can help identify problematic data sets.

SelectionFile type iconFile nameDescriptionSizeRevisionTimeUser
SelectionFile type iconFile nameDescriptionSizeRevisionTimeUser

  8k v. 2 Mar 23, 2015, 2:19 PM Fang-Cheng Yeh

Run this program to generate SIEMENS Free Model vector table  330k v. 3 Mar 23, 2015, 2:18 PM Fang-Cheng Yeh