[MINC-development] N3-1.12 changed behaviour using a mask

Wed Jun 15 10:08:21 EDT 2011

Hi Claude,

I have a follow-up question related to the lambda value. Do you know the 
formula that we can use to compute the lambda value for the current 
version of N3 which will give the same result as using 1.0e-7 in N3 1.10 
given the other settings?

Thanks,
Matthijs

On 11-06-14 11:32 PM, Matthijs van Eede wrote:
> Hi Claude,
>
>
> On 2011-06-14, at 6:25 PM, Claude LEPAGE wrote:
>
>> Hi Matthijs,
>>
>> I looked at the scan. This is a mouse. :-)
>>
> Yes :-) I don't work with anything else!
>
>> The mask you sent me covers the entire head. You would get better
>> results with a tighter mask that includes only the brain, since there
>> are lots of hyperintense voxels in the tissues surrounding the
>> brain. These can affect the splines in N3.
> I had no idea, thank you.
>
>> The old N3-1.10, which you used, had a major bug in it in terms of
>> damping. For a masked human brain in stereotaxic space at 1mm, the
>> damping was equivalent to about 3e-05 at a shrink factor of 4. (Note
>> that the old N3 was still good at those settings, so don't panic
>> about your old results.) This effective damping was normalized by
>> the number of sampling points in the mask in one hard-coded
>> coefficient. For a mouse with different voxel sizes, total brain size,
>> shrink factor, you need to be careful about the meaning of damping in
>> the old vs new N3. In the newer version, the formulation is now invariant
>> to sampling resolution and damping means damping. For example, resampling
>> at 0.5mm or 1.0mm will now give the same answer.
>>
> Good to know that the old results are still okay. I'll sit down with Jason and John, so they can tell me a bit more about damping and to figure out what settings are appropriate for mouse brains.
>
> Thanks very much!
> Matthijs
>
>> This is why you observed quite different results with 1.10 and 1.12.
>>
>> The instabilities in 1.11 or 1.12 were due to a too small damping
>> coefficient and likely also a too generous mask. You can also try
>> to add a bit of padding around the current field of view (use
>> autocrop -isoexpand).
>>
>> You should talk to Jason Lerch, also at Phenogenomics, for suitable
>> parameters for N3 (spline distance, for example). 8mm might be too
>> large a value for a mouse brain.
>>
>> Good luck,
>>
>> Claude
>>
>>>>> I have recently installed N3-1.12 together with ebtks 1.6.4, and I am
>>>>> getting different results when I run a non uniformity correction with a
>>>>> mask as compared to N3-1.11. In 1.12 the image intensities outside of
>>>>> the mask are "blown up"; they are about an order of magnitude brighter
>>>>> than the voxels within the mask. Are there changes between the two
>>>>> versions that can cause this?
>>>>>
>>>>> Thanks in advance,
>>>>> Matthijs
>>>> At first sight, this is unexpected. Does this happen on all subjects
>>>> or only on one of them? How large is your field of view outside the
>>>> mask? Does the mask touch the borders of the field of view? Is this
>>>> a regular t1 image? If not, what type of scan is it?
>>>>
>>> It happens in about 40% of the cases (about 20 files so far). The masks
>>> cover around 45% of the volume. I checked one of the files where this
>>> happens, the mask touches one of the borders of the volume. (Maybe
>>> interesting to note, is that in that area around the mask at that
>>> border, the output file is very bright) The files are t2 weighted. The
>>> resolution for these files are:
>>>
>>> image: unsigned short 0 to 65535
>>> image dimensions: zspace yspace xspace
>>>      dimension name         length         step        start
>>>      --------------         ------         ----        -----
>>>      zspace                    152        0.056         -4.2
>>>      yspace                    294        0.056        -8.19
>>>      xspace                    225        0.056        -6.27
>>>
>>>
>>>> In 1.12, I improved the convergence of the field outside the mask when
>>>> voxels are not 1mm. The number of iterations was tuned for 1mm voxels,
>>>> so 1.12 did not converge fully on much smaller voxel sizes and was
>>>> much slower (I'm working with histology data at 10 microns).
>>>>
>>>> What command did you use? You can try with slightly more damping
>>>> (default -lambda 1.0e-7 so try 1.0e-6).
>>> The commands I used are:
>>>
>>> nu_estimate -distance 8 -iterations 100 -stop 0.0001 -fwhm 0.15 -shrink
>>> 4 -mask mask.mnc input.mnc  output.imp
>>>
>>> and
>>>
>>> nu_evaluate  -mapping  output.imp  input.mnc input_nu_evaluated.mnc
>>>
>>> Using -lambda 1.0e-6 made it a lot better. It almost fixed it, but you
>>> can still see a slight "halo" on the outside of the corrected brain. I
>>> have uploaded the input and output files to McGill, and actually saw
>>> that the other version of N3 I was using was 1.10, not 1.11.
>>>
>>> Thanks very much!
>>> Matthijs
>>>
>>>> If you can send me a typical scan, you can dump it in:
>>>>      http://www.bic.mni.mcgill.ca/cgi-bin/BICupload
>>>> Provide the original scan, its mask, and the outputs you obtained with
>>>> 1.11 and 1.12, together with instructions on how to reproduce the
>>>> problem.
>>>>
>>>> Yours,
>>>>
>>>> Claude
>>>> _______________________________________________
>>>> MINC-development mailing list
>>>> MINC-development at bic.mni.mcgill.ca
>>>> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-development
>> _______________________________________________
>> MINC-development mailing list
>> MINC-development at bic.mni.mcgill.ca
>> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-development
> _______________________________________________
> MINC-development mailing list
> MINC-development at bic.mni.mcgill.ca
> http://www.bic.mni.mcgill.ca/mailman/listinfo/minc-development