Diffusion-weighted (DWI) imaging can provide insight into stroke. Computing tensors from DWI (Diffusion Tensor Imaging, DTI) can identify [damaged networks]. Diffusion measures such as kurtosis (DKI) can identify [microstructural changes]. Setting up a DWI scan provides many tradeoffs. For example, those seeking good DTI data tend to strive for high spatial resolution, while DKI requires multiple and strong b-values. Further, research teams working with healthy young adults utilize small high-density head coils with 32 or more channels. These coils allow faster acquisition (higher multi-band factors) and higher signal to noise. However, they are unsuitable for many clinical studies. Many individuals simply can not fit into these coils, as their heads are too big or they can not adjust their posture well (e.g. muscle rigidity after stroke). The dataset here seeks a balanced DWI scan that can be used for both DTI and DKI measures using the large Siemens 20-channel head(16)/neck(4) coil. We briefly describe the tradeoffs inherent with any diffusion scan, both to justify our choices and to help others when considering setting up a DWI scan.
Below we describe our choices. A general goal was to keep the echo time (TE) short. While a longer TE improves the T2 effects, very long TEs have poor signal to noise. This goal drove many decisions.
A major decision is the spatial resolution. For example, the Human Connectome Project acquires extremely high resolution isotropic 1.5mm voxels. In contrast, we selected 2.2mm voxels (i.e. three times the volume). Smaller voxels allow more spatial precision. However, they have substantially less signal to noise. The lower resolution allows faster EPI readout times with the associated benefits. So the goal was fewer, higher quality samples. Pilot work suggested this was required for reliable Kurtosis estimates.
Many DWI scans employ partial Fourier acquisition. Benefits include a tighter point spread function (PSF), the lower TE (increasing SNR), faster EPI readout (less spatial distortion). On the other hand, one has fewer sample (reducing SNR). Further, one can remove [Gibbs ringing artefacts] from the data. For these reasons, we chose not to evaluate partial Fourier, opting for full k-space coverage.
In-plane acceleration (SENSE/GRAPPA) is similar to partial Fourier: one acquires fewer samples of k-space with generally the same benefits and costs. However, since k-space is fully covered, one can still combat Gibbs ringing.
Multi-band can dramatically accelerate acquisition of EPI-based MRI data. For example, a multi-band factor of 2 acquires two slices of the brain simultaneously. This can allow shorter acquisitions (or more samples in the same amount of time). However, from first [principles one] one should be less agressive with multi-band when using a coil with fewer channels. Here we explored multi-band levels of 2 and 3. We used multi-band to reduce the repetition time (TR) with higher levels reducing acquisition time but expecting a bit less SNR (less T1 recovery). Our goal was to quantify this trade off. [Elsewhere] we keep the TR constant to evaluate the influence of multi-band on the 20-channel head coil.
Both SENSE and multi-band accelerate imaging by leveraging the different biases observed across different coils, one needs to be careful about combining these methods together. In particular, one needs to be concerned about aliasing artefacts when using high acceleration factors [fewer coils].
If one wants to avoid both partial Fourier and SENSE, one needs to worry about excessively long TEs. One needs to consider abandoning the [bipolar twice-refocused] sequneces for monopolar sequences. The reduces TE, but leads to more spatial distortion and more influence of background gradients. Hopefully, one can use modern tools like Eddy and TOPUP to undistort these images. We decided to evaluate this tradeoff by acquiring monopolar sequences without SENSE to a bipolar sequence with SENSE.
Finally, we acquired the shortest acquisition (MB=3) with the same participant. This allowed us to evaluate the cost of using the 20-channel head coil. While this coil is not suitable for our population, this provides a way to evaluate the consequences of using the larger head coil.
The data was acquired on a Siemens Prisma-Fit E11C using the product DWI sequences. Data is provided in DICOM format and can be converted to NIfTI format using [dcm2niix].
Nine series are provided. Each series provides 138 volumes (10 B=0, 64 B=1000, 64 B=2000) at the specified multi-band level plus a single-band reference image.
1. 20-Channel T1 scan (6:17)
2. 20-Channel DKI MB=2 monopolar TE=103 AP (10:36)
3. 20-Channel DKI MB=2 monopolar TE=103 RL (10:36) [this was planned as PA, but a bug in Siemens E11C copy references converted this to RL]
4. 20-Channel DKI MB=2 SENSE=2 bipolar TE=90 AP (10:14)
5. 20-Channel DKI MB=2 SENSE=2 bipolar TE=90 PA (10:14)
6. 20-Channel DKI MB=3 monopolar TE=107 AP (7:21)
7. 20-Channel DKI MB=3 monopolar TE=107 PA (7:21)
8. 20-Channel DKI MB=3 monopolar TE=107 AP (7:21)
9. 20-Channel DKI MB=3 monopolar TE=107 PA (7:21)
We used our [nii_preprocess] script to analyze the data. This uses a combination of FSL and MRTrix tools to denoise and undistort the dataset. Crucially, we used [dwidenoise] to measure the amount of noise, allowing us to estimate signal to noise.
The results are shown here:
![Signal to noise for each B-value]
In brief, the monopolar sequences exhibit better SNR than the bipolar/SENSE sequence. The 32-channel coil MB=3 sequence (7:21) has similar SNR to the 20-channel coil at MB=2 (10:36).
The image below compares the MB2 monopolar (left) with MB2 SENSE2 bipolar (right) image for one sampled B=1000 direction. The bipolar sequence does exhibit less spatial distortion, but has visibly poorer signal to noise.
![Monopolar (left) vs bipolar (right) note differences in spatial distortion and signal to noise (B=1000 scans shown)]