code written: 2020-06-21
last ran: 2020-06-23

Background

Failed segmentations. We performed tractography, with Slicer whitematteranalysis, across the entire SPINS sample. As shown in detail in a past report, we have not achieved satisfactory segmentation across all tracts, including tracts that should be well characterized by automated segmented (e.g., IOFF), and tracts of interest to our hypotheses (namely AF).

Scanner effect. We observed that the success of segmentation shows an effect of scanner. The visualization below shows number of tracts (out of a maximum possible of 72) segmented, across all n = 444 participants in the SPINS study with DWI data (CMH=137; CMP=32; MRC=67; MRP=70; ZHH=45; ZHP=93). On the left, we show differences in success by site; on the right, we show differences by scanner. It is clear that segmentation success is driven by scanner type: the n=4 Siemens scanners (CMP, MRP, and ZHP = Siemens Prisma; MRC = Siemens Tim Trio) are not segmenting as well as the n=2 GE scanners (CMH = GE Discovery; ZHH = GE Signa).

Present analysis

Purpose. Here, we closely review the outputs of each processing step, to ensure nothing has gone wrong. If it is determined that nothing is wrong with our processing steps(s) (code, input, output), we will continue to review the tractography, to understand at what step(s) are we losing key tract data.

Participants. We selected the same randomly-selected ‘good’ participant used in past analyses, as well as one ‘bad’ participant from each of the Siemens sites, defined by the lowest number of segmentations at each respective site. Thus, the participants included in the present analysis are as follows:

participant tracts (/72) scanner
comparator
CMH0135 68 GE
ZHH0030 68 GE
average segmentation
CMP0182 69 Siemens
MRC0053 62 Siemens
MRP0096 71 Siemens
ZHP0158 64 Siemens
worst segmentation
CMP0186 49 Siemens
MRC0061 51 Siemens
MRP0103 57 Siemens
ZHP0114 45 Siemens

[1] Raw data

Data parameters

There are small differences in parameters across sites, as well as hardware (head coil), but I believe these are all known and intended. The most notable differences is the PhaseEncodingDirection, which differences between the GE and Siemens sites, but this has been taken into account in Michael’s preprocessing steps.

comparator
average segmentation
worst segmentation
variable CMH0135 ZHH0030 CMP0182 MRC0053 MRP0096 ZHP0158 CMP0186 MRC0061 MRP0103 ZHP0114
SliceThickness 2 2 2 2 2 2 2 2 2 2
SpacingBetweenSlices 2 2 2 2 2 2 2 2 2 2
SAR 0.476 0.476 0.32 0.463 0.262 0.299 0.276 0.557 0.253 0.263
EchoTime 0.086 0.086 0.073 0.085 0.085 0.085 0.073 0.085 0.085 0.085
RepetitionTime 17 17 8.8 8.8 8.8 8.8 8.8 9.1 8.8 8.8
FlipAngle 90 90 90 90 90 90 90 90 90 90
Head coil channels 8 8 64 32 64 64 64 32 64 64
AcquisitionMatrixPE 128 128 128 128 128 128 128 128 128 128
ReconMatrixPE 128 128 128 128 128 128 128 128 128 128
TotalReadoutTime 0.085 0.085 0.036 0.044 0.033 0.036 0.036 0.044 0.045 0.033
PixelBandwidth 3906.25 3906.25 2300 2298 2300 2300 2300 2298 1630 2300
PhaseEncodingDirection j j j- j- j- j- j- j- j- j-

Directions

All participants have 60 directions, and they are spaced appropriately (note that the difference in colour and scale is a function if these images being generated with different software versions, and is not of consequence).

comparator
        CMH0135                  ZHH0030

average segmentation
          CMP0182                  MRC0053                   MRP0096                   ZHP0158

worst segmentation
          CMP0186                  MRC0061                   MRP0103                   ZHP0114

Data montage

The raw data look ok, and similar across participants with varying segmentation success.

comparator

average segmentation

worst segmentation

[2] Distortion correction

I think that there remain issues in distortion correction. In some cases, the images look more distorted after correction … ? This is especially so in the worst segmentation cases.

comparator

average segmentation

worst segmentation

[3] Tensor

Tensor fit with Slicer DWIToDTIEstimation. Note that the Slicer tutorial calls for the tensor to be fit before masking, which is evident here. The tensor fit is superior in GE than Siemens.

comparator

average segmentation

worst segmentation

Comparison to tensor fit with dtifit. For comparative purposes, we show the tensor fit with dtifit, on data that has not been preprocessed with BrainSuite. I think it appears superior.

comparator

average segmentation

worst segmentation

[4] Mask

Lastly, we look at the masks, made within Slicer, which we showed in a prior analysis to be sufficient for tractography. Again, we see that the quality of masking in superior at the GE vs. the Siemens sites.

comparator

average segmentation

worst segmentation