Print

---

Hi Matthew,

 

I have only been focusing on the restored images at the moment, as I have
been incorporating a bias correction step into the preprocessing of T1 and
FLAIR images.  I do think that multiple iterations of FAST give a more
homogeneous output image (with respect to bias).  The old version of FAST
often over corrected the bias, whereas FAST4 tends to under correct the
bias, however with subsequent iterations it usually tends toward an
(accurate) equilibrium.  We tend to get very good segmentation results
except in motor cortex and in the thalamus (these are non-human primate
scans), however the contrast in these regions is poor due to the large
amount of myelin present.  

 

Peace,


Matt.

 

  _____  

From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf
Of Matthew Webster
Sent: Monday, April 14, 2008 5:14 AM
To: [log in to unmask]
Subject: Re: [FSL] Fast4 Question

 

Hello Matt,                      

                       Thanks for the command line information. It's
possible that the differences you are seeing by rerunning with the restored
image is due to fast4's initial bias field estimation being different to
that used in the -I iterations. Hence by running twice you are getting the
initial correction applied twice ( on the original and restored images
respectively). I would be very interested if you think that running fast4
twice always gives a better segmentation or if it's data set dependent. The
--Hyper option is just a scaling factor ( set to < 0 to auto-estimate ) the
--lowpass option is in mm.
Many Regards
Matthew

 





Hi Matthew,

 

Thanks for your reply.  Here is an example commandline applied to a T1
image:

 

fast4 -b -o default T1Struct

fslmaths T1Struct -mul default_bias T1Struct_default

fast4 -b -I 8 -o I8 T1Struct

fslmaths T1Struct -mul I8_bias T1Struct_I8

fast4 -b -I 12 -o I12 T1Struct

fslmaths T1Struct -mul I12_bias T1Struct_I12

 

The files T1Struct_default T1Struct_I8 T1Struct_I12 are all the same.  If I
run the default, and then run the default again on the restored image I get
a different (somewhat better) result.  Also, what are the units of the
--lowpass and ---Hyper smoothing options?

 

Thanks,

 

Matt.

 

  _____  

From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf
Of Matthew Webster
Sent: Monday, April 07, 2008 6:39 AM
To: [log in to unmask]
Subject: Re: [FSL] Fast4 Question

 

Hi,

     We've done some tests with fast4 and we can't replicate what you've
seen with an invariant bias field for different --iter. Could you go into
more detail into the method you've used ( specific command lines etc)? Given
the smaller brain size you may want to lower the --lowpass and ---Hyper
smoothing options...

 

Many Regards

 

Matthew




I have found that running fast4, restoring the image with the bias field,
and then rerunning fast4 several times gives a better estimation of the bias
field than just running fast4 a single time.  I tried to duplicate this by
increasing the number of main-loop iterations during bias-field removal (the
--tier option) from 4 to 8, for example, but this gave exactly the same
result for the bias field restoration.  How would I get the same improvement
in bias field estimation without rerunning the program several times (and
wasting time calculating segmentations that I don't plan to use)?  Also, are
there any settings I should change for non-human primates with their smaller
brains?  It seems like sulci are often smoothed out of the segmented image.

 

Thanks,


Matt.