Thanks for the feedback.

I suspect that changing the warping regularisation would have the
greatest effect for crisper tissue probability maps.  The ones in SPM
were generated by affine registration, so are a little blurry.  If
they were sharper, the registration would benefit more from greater
flexibility.  The edges of the warped tissue probability maps should
then demarcate the brain more clearly.

It would probably also work a bit better with more knowledge about
tissues just outside the brain, which are often quite variable.
Segmentation struggles most with non-brain tissue close to the brain,
which has the same intensity as GM.

Best regards,
-John

On 22 July 2012 16:18, Lúcia Garrido <[log in to unmask]> wrote:
> Dear John,
>
> Many thanks for your responses and all the information. It's all very
> helpful.
>
> I varied the parameters you mentioned and, just in case this can be of
> interest to someone else, I summarize a few things below.
>
> 1. In my case, MRF seems to have worked better to clean up the segmentation.
> I used MRF of 2 as you suggested. I attach on example (example3_MRF2) and
> also how it compared with just the default setting of MRF = 0
> (example3_default_MRF2). Overall, as far as I could tell increasing the MRF
> to 2 removed better non-brain tissue, or at least gave it lower intensities,
> and didn't affect much the intensity values of actual grey matter. Although
> results from using MRF = 2 seemed better overall, results were still not
> perfect (see example4_MRF2).
>
> 2. Changing warping regularisation (I changed from 4 to 1) didn't seem to
> change things much. Is this expected?
>
> 3. Changing sampling distance from 3 to 1 (resolution in my scans is 1 mm)
> resulted in much longer segmentation (from about 10 mins to 2 hours and
> half) but not improved results for the specific problem I had. In fact, in a
> few cases, there's even more non-brain included.
>
> Many thanks and bets regards,
>
> Lucia
>
>
>
> On 19 July 2012 07:28, John Ashburner <[log in to unmask]> wrote:
>>
>> > I've been using new segmentation and had a few questions that I'd be
>> > very
>> > grateful if someone could answer or had any suggestions.
>> >
>> > I've done new segmentation and I think for many subjects it didn't work
>> > very
>> > well well. I've attached one example and there are many other subjects
>> > like
>> > this. It seems like the regions of the inferior temporal lobes were not
>> > segmented very well.
>>
>> They look mostly OK to me, although there is a bit of non-brain included.
>>
>> >
>> > I used the default options in new segment, so I used:
>> > - bias regularisation:  'very light regularisation (0.0001)
>> > - bias FWHM: 60mm cutoff
>> > - Used the tissue probability maps in TPM, 1 to 6, with these respective
>> > number of Gaussians: 2, 2, 2, 3, 4, 2.
>> > - MRF parameter: 0
>> > - Warping regularisation: 4
>> > - Sampling distance: 3
>> >
>> > So my questions are:
>> > 1. Do you have any suggestions about what I can do differently to get
>> > better
>> > segmentation?
>>
>> Segmentation largely relies on voxel intensities, so MRI sequences can
>> be optimised to make the GM intensity as different from that of other
>> tissues as possible.  This is probably not so helpful for the data you
>> have collected already though.
>>
>> Manual editing of the segmentation results would be another
>> possibility.  At the moment, manual segmentation is generally
>> considered to be as close to a gold standard as anything, so manual
>> editing should be considered acceptable - providing whoever is doing
>> the editing knows nothing about which subjects are in which group etc.
>>
>> There are also many other segmentation approaches that may be a better
>> model for your data.  You could try some of these.
>>
>>
>> >
>> > 2. Has anyone varied the parameters above systematically and would be
>> > willing to share what they found with the list?
>>
>> Try setting MRF parameter to 2, and see what happens.  This may clean
>> things up slightly.
>>
>>
>> > I know that any results are going to depend a lot on the data we have,
>> > but
>> > I'd appreciate any insights about the different parameters. The manual
>> > is
>> > very helpful in explaining the different parameters, but I'd like to get
>> > a
>> > better idea about the effects of each of them. Moreover, for certain
>> > parameters, I'm not sure even how to vary them. In particular, for the
>> > MRF
>> > parameter, warping regularisation, and sampling distance, I'm not sure
>> > within which intervals these parameters can vary...
>>
>> MRF parameter was a late addition to spm8.  I didn't want it to change
>> the output from the previous updates of spm8, so I gave it a default
>> value of 0.  A value of 2 may give slightly cleaner results, but if
>> you go much higher, you may begin to lose some of the detail from the
>> segmentation results.
>>
>> Warping regularisation can often be decreased slightly for T1w images.
>>  This controls the amount of freedom of the deformation part of the
>> model.  More regularisation gives less freedom.  I chose this for the
>> default setting of relatively heavy regularisation because with too
>> much freedom, the algorithm can begin to treat the liquid contents of
>> the eyeballs as CSF, and squash the eyes into the brain.  Really, I
>> should either include eyeballs as part of CSF tissue priors, or
>> introduce a new eyeball tissue class.  This would allow the
>> regularisation to be decreased in order to more closely match the
>> tissue priors with the scans.  The eyeball squashing problem is less
>> severe for T1w than it is for other contrasts (where CSF is more
>> clearly visible), so you may be able to decrease the amount of
>> regularisation by a factor of about 10.
>>
>> Sampling distance is set to a default value of 3mm.  If you have 1mm
>> resolution, you could get slightly better segmentations by decreasing
>> this distance to 1mm.  This is not the default because it would result
>> in the algorithm taking 27 times (3 x 3 x 3) as long to run.
>>
>> >
>> > 3. I also have a question about dartel, and I'm sorry if this sounds
>> > trivial... In the manual, it says that dartel takes about one week for
>> > 400
>> > subjects. But I've run dartel with around 200 subjects and it took about
>> > 8
>> > hours. I have an iMac with 2.8 GHz i7, and 16GB ram. It seems it should
>> > take
>> > longer than 8 hours, and was wondering if that estimation of a week was
>> > from
>> > a long time ago... I'm also happy that dartel takes more time if the
>> > registration results are better. So if someone has any advice about
>> > this,
>> > I'd really appreciate your help.
>>
>> Maybe I should update the manual.  The one week result was for my five
>> year old laptop.
>>
>> > The parameters I've been using for dartel are the default ones:
>> > - regularisation form: linear elastic energy
>> > - 6 outer iterations, each with 3 Gauss-Newton iterations, default
>> > regularisation parameters, time steps (varying from 1 to 64), and
>> > smoothing
>> > parameters
>> > - default optimisation settings
>> > Like for the segmentation case, I'd very much appreciate any information
>> > about the results of varying any of these parameters.
>>
>> The default settings work reasonably well for most data.  There are a
>> couple of settings for the linear elasticity.  The first one is the
>> amount of penalty on length changes, whereas the second one penalises
>> volume changes.  I haven't fully explored the effects of changing the
>> settings - although I'm pretty sure that there is some room for
>> improvement.
>>
>> Best regards,
>> -John
>
>
>
>
> --
> Lúcia Garrido
> Post-Doctoral researcher
> Vision Sciences Laboratory
> Harvard University
>
> [log in to unmask]
>
> http://sites.google.com/site/garridolucia/
>