ok, thanks, now make sense
On the other hand, residuals are:
res = data - pred
right?
2016-09-29 14:00 GMT+02:00 Saad Jbabdi <[log in to unmask]>:
> Hi - I know what ‘they’ are doing as I wrote that bit of code :)
>
> I just tried and I get identical results (except for a few voxels at the
> edge of the brain like I said). Here are the steps I suggested:
>
> In a terminal:
> dtigen -t dti_tensor.nii.gz --s0=dti_S0 -o dti_pred -r bvecs -b bvals -m
> nodif_brain_mask
>
> In matlab
> sse=read_avw('dti_sse');
> pred=read_avw('dti_pred');
> data=read_avw('data');
> mask=read_avw('nodif_brain_mask');
> mask=~~mask;
>
> data=reshape(data,numel(mask),size(data,4));
> pred=reshape(pred,numel(mask),size(pred,4));
>
> data=data(mask,:);
> pred=pred(mask,:);
>
> x=sum((log(data)-log(pred)).^2,2);
>
>
> figure,plot(x,sse(mask),'o')
>
>
> You should see that most voxels give identical results.
>
> Cheers
> Saad
>
>
>
>
>
>
>
>
> On 29 Sep 2016, at 11:59, Ana Maria Escorza <[log in to unmask]
> <[log in to unmask]>> wrote:
>
> It remains different from that obtained with dtifit. The values are higher
> with that formula
> Maybe they do something like this:
>
> http://stats.stackexchange.com/questions/16845/log-squared-error-accuracy-
> justification-data-mining-competition
>
> What do you think?
>
> 2016-09-29 12:52 GMT+02:00 Saad Jbabdi <[log in to unmask]>:
>
>> It’s good enough for most uses. Also you probably only have negative
>> values at the edge of the brain probably caused by spline interpolation
>> during resampling.
>>
>> Cheers
>> Saad
>>
>>
>>
>>
>> On 29 Sep 2016, at 11:50, Ana Maria Escorza <[log in to unmask]
>> <[log in to unmask]>> wrote:
>>
>> but is it correct ignore values negatives in order to obtain sse?
>>
>> 2016-09-29 12:47 GMT+02:00 Saad Jbabdi <[log in to unmask]>:
>>
>>> data(data<0)=0
>>>
>>>
>>>
>>>
>>>
>>> On 29 Sep 2016, at 11:41, Ana Maria Escorza <[log in to unmask]
>>> <[log in to unmask]>> wrote:
>>>
>>> Thanks, that is for 0 values.
>>>
>>> But my sse has complex values after that. And it is different from sse
>>> obtain from dtifii
>>>
>>>
>>> Cheers,
>>> Ana.
>>>
>>> 2016-09-29 12:35 GMT+02:00 Saad Jbabdi <[log in to unmask]>:
>>>
>>>> An easy way to do this in matlab:
>>>>
>>>> SSE = sum ( (log(data+~data) - log(pred+~pred)).^2, 4 )
>>>>
>>>> The trick of doing x+~x means when x=0, then x+~x=1 and the log is
>>>> zero, so doesn’t count in the sum.
>>>>
>>>> Cheers
>>>> Saad
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> On 29 Sep 2016, at 11:08, Ana Maria Escorza <
>>>> [log in to unmask] <[log in to unmask]>> wrote:
>>>>
>>>> You mean do not do the logarithm of negative values?
>>>>
>>>> Remove negative values and put in after do the log?
>>>>
>>>> 2016-09-29 12:02 GMT+02:00 Saad Jbabdi <[log in to unmask]>:
>>>>
>>>>> In matlab, log is ln
>>>>>
>>>>> Just remove negative values before you calculate the sse
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On 29 Sep 2016, at 10:49, Ana Maria Escorza <
>>>>> [log in to unmask] <[log in to unmask]>> wrote:
>>>>>
>>>>> Hi,
>>>>>
>>>>> with log you mean log10 or ln?
>>>>>
>>>>> what you mean with safeguards?
>>>>>
>>>>> I do that I obtain complex numbers in sse.
>>>>>
>>>>> Cheers
>>>>>
>>>>> 2016-09-29 10:51 GMT+02:00 Saad Jbabdi <[log in to unmask]>:
>>>>>
>>>>>> Hi - I believe DTIFIT outputs the sum squared error in log-space,
>>>>>> i.e. sum( { log(data) - log(prediction) }^2 )
>>>>>>
>>>>>> There are also a couple of safeguards inside DTIFIT to make sure it
>>>>>> can take the log.
>>>>>>
>>>>>> Cheers
>>>>>> Saad
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 29 Sep 2016, at 08:21, Ana Maria Escorza <
>>>>>> [log in to unmask] <[log in to unmask]>> wrote:
>>>>>>
>>>>>> Once I have got the original DWIs and 'DWIs from DTs' exported
>>>>>> (with dtigen) and load into matlab, I try to calculate the sse of tensor
>>>>>> estimation that is calculated by sum((orDWI-DWI_DT).^2,4)
>>>>>>
>>>>>> But I didn't obtain the same that with --sse in dtifit, why?
>>>>>>
>>>>>> 2016-09-28 18:22 GMT+02:00 Ana Maria Escorza <
>>>>>> [log in to unmask]>:
>>>>>>
>>>>>>> How Can I subtract? with fslmaths from both .nii?
>>>>>>>
>>>>>>> 2016-09-28 18:18 GMT+02:00 Saad Jbabdi <[log in to unmask]>:
>>>>>>>
>>>>>>>> Hi - you can use ‘dtigen’ to generate predictions given the DTI
>>>>>>>> tensor (which you can save in DTIFIT with —save_tensor) and then substract
>>>>>>>> that prediction from the data.
>>>>>>>>
>>>>>>>> Cheers
>>>>>>>> Saad
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On 28 Sep 2016, at 14:42, Ana E. <[log in to unmask]
>>>>>>>> <[log in to unmask]>> wrote:
>>>>>>>>
>>>>>>>> Hello,
>>>>>>>>
>>>>>>>> I am looking for a function to obtain the residual DWI from the
>>>>>>>> diffusion tensor.
>>>>>>>>
>>>>>>>> I see that I can use --sse in dtifit, but this is for 3D volume. I
>>>>>>>> would like 4D.
>>>>>>>>
>>>>>>>> Thanks
>>>>>>>>
>>>>>>>> Ana.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>
>>>
>>
>>
>
>
