[Simnibs-discuss] Charm Segmentation Issues in Simnibs 4 Beta

Oula oulap at drcmr.dk
Thu Aug 11 08:19:46 CEST 2022


Hi Joel,

I'd be interested to look at cases 1) and 2). I didn't have any subjects 
with severe atrophy to test on so I think these can certainly be 
improved. Also having the cortical GM pushed all the way up to the CSF 
surface sounds strange.

W.r.t. fixing the affine: there are examples on how to do this in the 
documentation. If you click through tutorial -> Creating Head Models and 
then go all the way down on that page and click on "Fixing the Affine 
Registration". This should give you some pointers on how to do that. 
I'll also implement a flag for passing a registration matrix and 
skipping the optimization.

Best,

Oula


On 8/10/22 17:56, Joel Upston wrote:
> Oula,
>
> Thank you for the explanation and I found as you said that “mni” is 
> faster but does seem to fail more than “atlas”. However  after more 
> testing with these settings it looks like it is still having issues 
> about 40% of the time, this is over about 20 subjects I have tested 
> some with T1 and T2 others with just T1. Some of the issues are 
> summarized as follows:
>
> 1) The template appears to be registered correctly and the tissue 
> types are correctly segmented except the gray matter which as I 
> mentioned before is “pushed” right to the edge of the csf inner layer. 
> In these cases running the same scan through headreco with cat12 seems 
> to generate the correct gm surface but then have to use outside 
> programs to combine the two segmentations via editing, use the 
> add_tissues_to_upsampling and then rerun the meshing.
>
> 2) Some of the population of this dataset is above 50, with some 
> atrophy and large ventricles. The ventricles that are often large, 
> fail in the charm segmentation. My solution is to again use the 
> headreco segmentation ventricles in the charm segmentation.
>
> 3) The probabilistic atlas is still not in registration with the T1, 
> or could be improved. Would it be possible to add in a 12 parameter 
> registration matrix (similar to the init transform) that it is forced 
> to used, that way if another outside program such as AFNI or spm gets 
> a better registration I can use that? I have noticed that using the 
> init transform with the proper alignment ends up finding the same 
> incorrect minimum in the registration optimization.
>
> Overall I am happy to share some of this data as well to help with 
> some of these issues. Thank you for all your help and all your 
> contributions to automate a more accurate head model for the field.
>
>
> Thank you,
>
> Joel Upston
>
>> On Aug 8, 2022, at 12:30 AM, Oula <oulap at drcmr.dk> wrote:
>>
>> 
>>
>> *  [EXTERNAL]*
>>
>> Hi Joel,
>>
>> The standard init_type ("mni") does a registration between the T1w 
>> MNI template and the input scan using a correlation metric. This 
>> assumes that the tissue intensities of the input scan and the MNI 
>> template are more or less the same, and it can sometimes fail if 
>> that's not the case (even if the input is a T1w scan).
>>
>> Init_type="atlas" does a registration between the input and a 
>> probabilistic atlas using a mutual information-based metric (similar 
>> to what SPM12 does). This metric doesn't care about the specific 
>> tissue intensities of the input, so you could, for example, feed in a 
>> T2w scan and it should still work.
>>
>> The reason I set "mni" as default is that it's faster and seemed to 
>> work well on the cases I tested it. I'll try to do more testing on 
>> different T1w scans to see if we should set the "atlas"-option as 
>> default.
>>
>> Best,
>>
>> Oula
>>
>>
>> On 8/5/22 18:12, Joel Upston wrote:
>>>  Oula,
>>>
>>> Thank you for the quick response. Yes the template alignment seems 
>>> to be a bit better from the affine from headreco, than that of what 
>>> is given through the charm optimization. I tried the noneck which 
>>> didn’t change it but just tried the “atlas” init_type parameter and 
>>> that worked. How exactly is it different?
>>>
>>> Thank you,
>>>
>>> Joel Upston
>>>
>>>> On Aug 5, 2022, at 5:18 AM, Oula <oulap at drcmr.dk> wrote:
>>>>
>>>> 
>>>>
>>>> *[EXTERNAL]*
>>>>
>>>> Hi Joel,
>>>>
>>>> Did you check the template alignment after forcing the affine from 
>>>> headreco? You can also try switching the init_type from "mni" to 
>>>> "atlas" in the [samseg] section of the settings file. It also looks 
>>>> like the scans don't have neck so please include the --noneck flag 
>>>> if that's the case.
>>>>
>>>> If you've tried all these things already, I'm happy to have a look 
>>>> if you're allowed to share a scan or two from the data set.
>>>>
>>>> Best,
>>>> Oula
>>>>
>>>>
>>>> On 8/5/22 12:32, Joel Upston wrote:
>>>>> 
>>>>> Hi there,
>>>>>
>>>>> I have worked with the Simnibs 3.2.6 headreco to do segmentation 
>>>>> and now trying out the charm segmentation routine in the Simnibs 4 
>>>>> Beta release. I have multiple subjects (about 10 subjects tried so 
>>>>> far) that have looked good in the headreco segmentation in terms 
>>>>> of the brain segmentation especially using the T1 and T2 but when 
>>>>> I try this with the charm segmentation, I run into a problem about 
>>>>> 50% of the time. The problem is that the cortical surface is 
>>>>> smooth and pushed up right next to the csf boundary layer, but the 
>>>>> wm surface looks good. I have attached an image of what is 
>>>>> happening to show a better description. I have made sure that the 
>>>>> template is aligned by forcing in the code to use the Simnibs 3 
>>>>> warp from MNI2Conform 12 parameter alignment, where the previous 
>>>>> headmodel seemed to be  accurate. I also try to change some of the 
>>>>> settings in the settings.ini file such as all parameters in the 
>>>>> segment section and th_initial in the surface section with no 
>>>>> avail. As I run the same subject in the Simnibs 3.2.6 the 
>>>>> headmodel looks good, I have tried to edit via switching out the 
>>>>> brain+csf segmentation from simnibs3 to simnibs 4 and reran the 
>>>>> mesh after editing. This works reasonably well but the some of the 
>>>>> external to brain structures are affected, plus requires double 
>>>>> the runtime to run versions 3 and 4. Have you noticed a problem 
>>>>> similar to this and do you have any recommendations to try? Thank 
>>>>> you for all your work and I appreciate the research to keep 
>>>>> improving this modeling.
>>>>>
>>>>> Thank you,
>>>>>
>>>>> Joel Upston
>>>>>
>>>>> Correct_GM_surface_from_headreco.png
>>>>>
>>>>> Incorrect_GM_surface_from_charm.png
>>>>> Screenshot%20from%202022-08-04%2012-51-01.png
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.drcmr.dk/pipermail/simnibs-discuss/attachments/20220811/74195ad1/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Correct_GM_surface_from_headreco.png
Type: image/png
Size: 128092 bytes
Desc: not available
URL: <https://mailman.drcmr.dk/pipermail/simnibs-discuss/attachments/20220811/74195ad1/attachment-0003.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Incorrect_GM_surface_from_charm.png
Type: image/png
Size: 104518 bytes
Desc: not available
URL: <https://mailman.drcmr.dk/pipermail/simnibs-discuss/attachments/20220811/74195ad1/attachment-0004.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Screenshot%20from%202022-08-04%2012-51-01.png
Type: image/png
Size: 370088 bytes
Desc: not available
URL: <https://mailman.drcmr.dk/pipermail/simnibs-discuss/attachments/20220811/74195ad1/attachment-0005.png>


More information about the Simnibs-discuss mailing list