Thanks to everyone who's replied so far. I agree about the tradeoff,
at the point where 1% are going wrong if you change the algorithm
you'll often find a different 1% decide to go astray. Generally I'm
happier (because it's more easy to incorporate as a workflow and
simpler to show you are being consistent) to have checking followed up
by a second strategy for failures. Obviously with the ADNI T1 there's
the luxury of being able to do a few things that rely on it really
being T1. I'll have a look at the auto_reorient script some people
have mentioned, have had a little success with quickly trying a
similar approach based on initially doing coregister to MNI (using
normalised cross correlation).
If anyone is interested, I've put two nifti on our dropbox, one
improves with improved initial corregistration, the other doesn't
really, there's some spillage into dura of GM, WM and CSF classes.
I'll have to send the links privately on request to keep within the
ADNI data sharing policy.
Best wishes,
Ian
On 29 April 2015 at 12:06, John Ashburner <[log in to unmask]> wrote:
> Usually, the AC is not particularly close to the centre of the magnetic field. I guess I could shift the default starting estimates so that they better reflect how subjects are typically positioned in the scanner. This might be a bit messy though, and would require a bit of user re-education.
>
> Alternatively, I could have used a two-pass procedure that re-ran the affine registration for all subjects, using starting estimates based on a mean or median of the first pass solutions. Many would not like this approach because they want the same answer irrespective of which scans were processed together.
>
> There's always a trade-off about how general to make things. For example, I could make the affine registration more robust for T1w data, at the expense of it failing badly when applied to other image contrasts. Simple things, such as automatically moving the centre of mass of the images would work for some data, but not others. If an algorithm has to handle both T1 maps as well as R1 maps (where T1 = 1/R1), then assumptions about background signal being close to zero can not be made.
>
> Bottom line: fix the headers and things will work better. This fix does not need to be especially accurate, so I doubt that the twiddling of different users will result in systematically different segmentations.
>
> Best regards,
> John
>
>
> Ian Malone <[log in to unmask]> wrote:
>>Thanks, I've occasionally used similar tricks in the past, but of
>>course it leads to concerns about reproducibility if you've got a
>>person twiddling things, so I'm hoping there's something that can be
>>done more reliably. I can try an alternative registration strategy to
>>MNI for example.
>>
>>On 28 April 2015 at 16:07, MCLAREN, Donald <[log in to unmask]> wrote:
>>> Ian,
>>>
>>> Use the Display tool to reorient the images to match the origin and rotation
>>> of the MNI template. Then the segmentation should go smoothly.
>>>
>>>
>>>
>>
>>> On Tue, Apr 28, 2015 at 9:35 AM, Ian Malone <[log in to unmask]> wrote:
>>>>
>>>> Hi,
>>>>
>>>> I've been running SPM12 on the accelerated T1 scans from ADNI-2
>>>> (Alzheimer's Disease Neuroimaging Initiative), and found a few (about
>>>> 5 out of 700) images where the segmentation is poor (e.g. attached).
>>>> In most of those cases it looks like there may be a misregistration of
>>>> the underlying tissue prior maps, e.g. extents of csf, wm and gm seem
>>>> shifted in one direction. Loading the mwc classes together with the
>>>> TPM in check reg for this one shows it to be misaligned. Is there any
>>>> advice for dealing with this issue when it arises?
>>>> This being ADNI data there shouldn't be a problem sharing it with
>>>> anyone who wants to take a look.
>>>>
>>
>>--
>>imalone
>>http://ibmalone.blogspot.co.uk
--
imalone
|