Hello,
the formula SNR = meanint/std(noise) is the correct one. The SNR in
the noise-file is 2 times weaker than the one you want originally due
to that error in the script I mentioned in the previous e-mail.
Until I fix this, the best is to specify sigma instead of SNR when
you are running POSSUM (sigma = 64*required_sigma - this is due to
POSSUM processing which I explained in the previous e-mail). You can
calculate the required_sigma as meanint/SNR. The meanint calculate
from the generated image to which no noise is added.
The output image will have std(noise)=required_sigma and SNR=meanint/
required_sigma. Ignore the noise-file until it is fixed.
Best,
Ivana
On 14 Oct 2008, at 09:18, Bjorn Roelstraete wrote:i
> Dear Ivana,
>
> Thank you for clearing that out. However, if i use my numbers (I
> get them from the noise-file that is created after running possum)
> with your formula, it looks like SNR = meanint/128*sigma instead of
> SNR = meanint/64*sigma. For example 10 = 33.58/128*.026236 and 20 =
> 33.58/128*.013118 and so on. On the other hand, if i subtract the
> noise image from the no_noise image and calculate the std, I do get
> 64*.026236= 1.67. So it looks as if the formula would have to be
> SNR = meanint/2*std(noise).
>
> Bjorn
>
> Ivana Drobnjak wrote:
>> Dear Bjorn,
>> Sigma is the standard deviation of the noise (std(noise)), SNR is
>> SNR=meanint/std(noise) . meanint is calculated over the brain
>> voxels and ingnoring the background.
>> However, I think you have spotted an inaccuracy in POSSUM GUI
>> and scripts regarding the noise specification and SNR. Sigma in
>> the GUI is 64 times (dimX) smaller than the sigma of the output.
>> This extra number is due to the fact that the noise is originally
>> added to the raw signal which is than converted into the image
>> using FFT and that scalar is needed. So sigma that you specify in
>> the GUI (or in case you specify the SNR) does not correspond to
>> the actual sigma of the output (which you should by the way
>> calculate using fslstats -s option from the image you obtain by
>> subtracting the noise and nonoise image).
>> I will fix this asap so that the terminology is consistent.
>> Meantime, the best is not to use the SNR option but to specify
>> 64*required_sigma when you are running POSSUM. You can calculate
>> the required_sigma as meanint/SNR. The meanint calculate from the
>> generated image to which no noise is added.
>> Hope this helps,
>> Ivana
>> On 13 Oct 2008, at 13:21, Bjorn Roelstraete wrote:
>>> Dear list,
>>>
>>> In Possum there are 2 ways of specifying the noise level. You can
>>> add thermal noise with a certain SNR, or you can specify a value
>>> for sigma. Can someone explain the relationship between SNR,
>>> sigma, mean intensity, etc?
>>> In my case, with a specified SNR of 20, i get a sigma of .
>>> 01311808984375000000 with a medintensity of 33.582310. For a SNR
>>> of 10, i get a sigma of .02623617968750000000 with a medintensity
>>> of 33.582310. When i calculate the standarddeviation of the
>>> image, i get 5.01, so clearly sigma != stdv. What is the
>>> difference between them? How are they related? And most
>>> importantly, how are they calculated?
>>>
>>> The literature is very vague on this issue. For example, in one
>>> article i found that SNR = (meanintensity activation -
>>> meanintensity rest)/standarderror noise. Is this correct? If so,
>>> is this on the voxel, ROI or image level? Also, what is the
>>> relation between my standard deviation (as calculated with
>>> fslstats), the standarderror of the noise (whatever that is) and
>>> sigma as specified in Possum?
>>>
>>> Many thanks,
>>>
>>> --
>>> Bjorn Roelstraete
>>> Ghent University
>>> Department of Data-analysis
>>> H. Dunantlaan 1, B-9000 Gent, Belgium
>>> Tel: 32-9-2646434
>>> E-mail: [log in to unmask]
>>>
>
> --
> Bjorn Roelstraete
> Ghent University
> Department of Data-analysis
> H. Dunantlaan 1, B-9000 Gent, Belgium
> Tel: 32-9-2646434
> E-mail: [log in to unmask]
>
|