thanks Mark.
I went back and used fast 4.1 to segment the data I had previously
seg'd using fast 3, to compare versions head to head. here's my
conclusions, posted here in case anyone else cares about this issue
down the road.
conclusions: there are strong similarities and differences between
versions, in terms of results.
- the fast 3 results are effectively the same as fast 4.1 if you only
care about within-study variation. fast 3 results correlate very
strongly with fast 4.1 results (r's > .98). so, if you have healthy
subjects and only care about relative effects, the version of fast
probably does not matter.
- fast 4.1 gives me systematically smaller estimates for grey matter,
larger estimates for white matter and CSF, and smaller G / W ratio. I
see this pattern in every subject, for 121 subjects (= a mix of male
and female healthy adults, age 18-40.) if you care about actual
volume estimates (like how many cc's is a given human's brain), this
will matter.
- if anything the 4.1 grey and white estimates are probably more
accurate (at least in relative terms), based on the wider literature.
I was thinking of G/W ratio being ~1.6 extrapolating from Zhang &
Sejknowski 2000 PNAS (their fig 2), but other reports put the G/W
ratio at 1.35 for women and 1.27 for men (= Allen et al 2003
NeuroImage). using fast 4.1 I get a G/W ratio of 1.30 in 121 subjects
(mixed male and female). my other dataset with G/W ratio of 1.14 was
only men, so maybe the 1.14 ratio is not so bad (still seems a tad
low versus 1.27, but whatever, maybe scanner effects are possible too).
--Jeremy
gory details:
starting from T1 images in 121 subjects, I get a correlation of .986
(gray) and .987 (white) for volume when processed using fast 3.x
versus fast 4.1 (i.e., correlating estimates from f3 with f4). the
fast 4.1 estimates are systematically smaller in terms of gray volume
(83 - 147 cc smaller = smaller in every single subject, mean 110 cc
smaller). conversely, white matter is systematically larger when
estimated by fast 4.1, by 37 to 83 cc (larger in every single
subject, mean 61 cc larger).
so overall brain volume (gray + white) and the gray / white ratio is
less, in every single subject, when using fast 4.1. the G / W ratio
drops from 1.67 (fast 3.x) to 1.30 (fast 4.1). (in my other data set
the ratio was 1.14 in men only using fast 4, so some of the
differences in the G/W ratio appears to be the sample / scanner, but
there's still something else systematic going on with fast.)
> What do the differences between the pve images look like between
> the methods?
> That is, take the grey matter PVE in both cases and do the
> subtraction of them,
> then view this in FSLView. Does it look unusual? Does the
> difference tend to
> concentrate in some areas (e.g. deep-grey structures, border of GM
> and CSF, etc.)?
> Or is there just a general bias one way across the whole image?
I don't have pve images from the FAST 3 analyses, but do still have
the segmentation classification images with all three tissue types.
from these I isolated gray matter using fslmath functions, and then
subtracted the fast 4 gray matter pve image from the fast 3 gray
matter image. I only did this for one subject. I am not sure what to
look for exactly, but it does not look obviously odd: lots of small
discrepancies spread all over the brain, both positive and negative
differences, and zero in places that should be white matter or
ventricles.
> On 21 Jun 2009, at 18:53, Jeremy Gray wrote:
>
>> Hi all,
>>
>> I have a question about how FAST 4.1 differs from earlier
>> versions, esp 3.x. I get quite different results in two large
>> samples, in terms of grey matter to white matter ratio (ratio =
>> 1.14 from FAST 4.1, vs ratio = 1.67 from 3.x). This suggests big
>> difference(s) in classification between versions. there are many
>> differences between versions, obviously, but I did not expect
>> these to impact the ratio much if at all. I think a ratio of 1.6
>> is much more in line with the literature.
>>
>> any ideas what's going on, or how to track it down? I searched the
>> archives but did not see anything relevant, sorry if I missed it.
>>
>> thanks in advance,
>>
>> --Jeremy
>>
>> gory details:
>> I ran T1 images from 114 subjects through bet (and checked that
>> they look reasonable after bet). I then used FAST 4.1 (on CentOS
>> 5.3, 64-bit) to segment them using defaults:
>>
>> fast -t 1 -o <image> <image>
>>
>> I looked at some of the resulting pve_* images and they look
>> reasonable in fslview. from the pve's I get CSF, grey, and white
>> matter volumes in mm^3 using fslstats and bc, as described on the
>> FAST web page (http://www.fmrib.ox.ac.uk/fsl/fast4/index.html)
>>
>> my intuition is that the grey to white matter ratio should be
>> pretty robust against scanner and sample differences (but maybe
>> that's wrong). in the sample I'm using FAST 4.1 for, the ratio of
>> GM : WM is 1.14. This is markedly lower that the G/W ratio of 1.67
>> I got in another large sample (different scanner & subjects, but I
>> don't think that should have a huge impact on the ratio). this was
>> using FAST 3.x (again on T1 images run through bet). so think that
>> the version of FAST is where the difference is, and that in my
>> hands 4.1 is not doing what it should.
>>
>> also, there seems to be more CSF when using FAST 4.1 (~25% of
>> intra-skull volume, vs ~15% with 3.x).
>>
>> any ideas on how to best track down what is going on (esp with 4.1)?
>>
|