[Camino-users] Voxel Classify - Human Connectome Project Data

[Philip A Cook]

I often attempt to clear up the isolated voxels by removing small clusters. It takes a few steps to do this and reconstruct the corrected SH map, but it allows me to set the threshold a bit higher and then deal with some of the false positives after the fact.

I agree with Danny that fitting some order 4 unnecessarily is not usually too bad, as long as you aren't doing something like restricting the compartment size of the tensors that might force a badly incorrect solution.


On May 8, 2013, at 1:31 AM, Alexander, Daniel wrote:

> It doesn't look great, but I can believe the result: the dense regions of order 4 are in the right places.  Did you account for the spatially varying b-values in the HCP data?  I guess not and that may be the cause some of the speckle, but this classification procedure is fairly crude anyway.  You should err on the side of too many order 4, which it looks like you have done, as fitting the two-tensor model where one would do is not too detrimental.  Worth continuing I think.
> 
> Danny
> 
> 
> On 6 May 2013, at 12:47, Niklas Kasenburg <niklas.kasenburg at diku.dk>
> wrote:
> 
>> Hi,
>> 
>> thanks for the quick reply. I was aware of the fact that I can change the thresholds and I also did this to then compute the real classification with integer values. But those results looked even worse to show you were my problem is.
>> 
>> So I tried the suggestion from Danny to use only one of the shells. I tried this with all the three different b-values by extracting the corresponding images and b-vectors in addition to the b=0 values.
>> Though the results get a little bit better, I think they are not as they should be. I attached an image of the classification results from using only b=1000 and as thresholds: 1.0E-50 1.0E-08 1.0E-03. Yellow voxels are order 4 and orange/red ones are order 2. As you can see most of the white matter voxels are classified as order 4. Do you think the classification is reasonable enough to continue computing the tensors? Do you have any other suggestions of how to improve the classification? Any suggestions are welcome.
>> 
>> Best,
>> Niklas
>> 
>> Am 30.04.2013 18:12, schrieb Alexander, Daniel:
>>> Phil is right about the choice of thresholds, but I think there may be an additional issue here to do with the data.  The HCP data has three HARDI shells with different b-values whereas the voxelclassify algorithm expects a single shell with one b-value and is likely to overestimate the complexity if you give it all three shells.  Probably the best way to get around this would be to give voxelclassify just the high b-value shell.  The subsequent multi-tensor fitting parts should work fine with all three shells together, but using the classification from just the high b-value shell.
>>> 
>>> Danny
>>> 
>>> 
>>> On 30 Apr 2013, at 16:41, Philip A Cook <cookpa at mail.med.upenn.edu>
>>> wrote:
>>> 
>>>> Hi,
>>>> 
>>>> You can use the sliders at the bottom to adjust the thresholds for each classification. The appropriate thresholds will likely be different for data with different schemes or scanners.
>>>> 
>>>> Air voxels contain a highly variable signal that is unlikely to look anything like Gaussian diffusion, so it is not surprising that the hypothesis of SH order 0 is rejected in those voxels. You should be able to get rid of them with a brain mask.
>>>> 
>>>> On Apr 30, 2013, at 7:48 AM, Niklas Kasenburg wrote:
>>>> 
>>>>> Hi,
>>>>> 
>>>>> currently I try to reconstruct a multi tensor diffusion model. The data I am working with is publicly available (http://www.humanconnectome.org/). More specific: the processed diffusion data of subject  100307. I tried to follow the Camino tutorial "multi-fibre/HARDI Reconstruction" using this data. The problem occurred already in the voxel classification step. I used the following command line:
>>>>> 
>>>>> voxelclassify -inputfile dwi.Bfloat -bgthresh 600 -schemefile schemefile.scheme -order 4 > dwi_VC.Bdouble
>>>>> 
>>>>> where dwi.Bfloat is the with image2voxel transformed diffusion data and schemefile.scheme is the schemefile. When looking at the result with:
>>>>> 
>>>>> vcthreshselect -inputfile dwi_VC.Bdouble -datadims 144 168 110 -order 4
>>>>> 
>>>>> a lot of background voxels are classified as 2nd or 4th order. In addition, with the preset thresholds there are more 4th order than 2nd order voxels (see attached screenshot). So all in all the classification looks somehow wrong for me. Any ideas what went wrong or what I should do differently? Any help is much appreciated.
>>>>> 
>>>>> By the way: I followed the tutorial using the tutorial data and everything worked fine.
>>>>> 
>>>>> Best regards,
>>>>> Niklas Kasenburg, PhD student
>>>>> University of Copenhagen
>>>>> Department of Computer Science
>>>>> 
>>>>> <VC_result.png>_______________________________________________
>>>>> Camino-users mailing list
>>>>> Camino-users at www.nitrc.org
>>>>> http://www.nitrc.org/mailman/listinfo/camino-users
>>>> _______________________________________________
>>>> Camino-users mailing list
>>>> Camino-users at www.nitrc.org
>>>> http://www.nitrc.org/mailman/listinfo/camino-users
>>>> 
>>> 
>> 
>> <voxel_classify.png>_______________________________________________
>> Camino-users mailing list
>> Camino-users at www.nitrc.org
>> http://www.nitrc.org/mailman/listinfo/camino-users
> 
> 
> _______________________________________________
> Camino-users mailing list
> Camino-users at www.nitrc.org
> http://www.nitrc.org/mailman/listinfo/camino-users