 |
 |
Hi, This appears to be a bug in the multi-channel bias correction method implemented in FAST. We will try to fix this for future releases. In the meantime (and this is what we always do internally, and why we have not spotted this before) you should bias correct all of your individual images using FAST (single channel), then register them, and then use multi-channel FAST. The segmentation part of multi-channel FAST works fine - it is only the bias field correction that is poor, so if you fix the bias field initially then everything should work well. All the best, Mark > On 20 Oct 2017, at 11:52, Ami Tsuchida <[log in to unmask]> wrote: > > Hello, > > I was comparing the performance of the tissue segmentations using the following three; > > -SPM New Segment using T1w and T2Flair > -FAST using T1 alone > -Multi-channel FAST using T1w and T2Flair > > SPM used the T1 and Flair images with skull, TPM specifically made for this cohort, and classified tissues into 6 classes. > For FAST, I performed BET on T2 Flair, used the mask to extract brain from T1 as well, and provided extracted T1 and Flair brains as input. > > I was expecting them to perform more or less equally well, but to my surprise, Multi-channel FAST deteriorated the segmentation. It classified dorsal and anterior cortex region as one tissue type, and white matter and posterior cortical region as another tissue type. > > I tried a few different things, and found that using already bias-field corrected T1 and T2Flair (output of SPM segmentation) mitigate the problem. Or, if I provide the tissue priors for the cohort, it produces more reasonable-looking segmentations. > > I'm a bit confused since first, I thought that FAST implements its own BF correction, and the bias in the original input does not seem so strong as far as I can see from viewing images, so I'm not sure why it's having such a huge impact. Also, the single-channel FAST uses non-BF corrected T1 and performs perfectly well. I also tried single-channel FAST with non-BF corrected T2 Flair and this also actually works reasonably well. Why does adding the two together deteriorates the performance? > > Am I missing something more fundamental? > > I use FSL5.0, and was originally running it through nipype, but tested it again using command line as follows; > > fast -o out_name -g -S 1 T1 > > or for multichannel > > fast -o out_name -g -S 2 T1 T2flair > > Thank you in advance for any input! > > Ami > >