[Mristudio-users] DTI method: VBA, MRIstudio vs. TBSS

Sat May 15 12:46:54 EDT 2010

Dear Susumu Mori,

Sorry that I misunderstood some of your previous ellaboration. Now I get
closer to your points, as well as the helpful arguements :) I  guess I may
add some points to my defense for VBA.
Thanks again very much for your further explanation!

Best wishes,
Gao

On 15 May 2010 06:24, susumu mori <susumu at mri.jhu.edu> wrote:

> Well, I didn't mean TBSS was better than the other. Different granularity
> means different sensitivity/specificity to different types of diseases.
>
> For example, when we are studying brain growth and tissue atrophy in
> certain neurodegenerative disease, it happens in a global manner. There is
> an overall tendency like frontal lobe develop earlier than the occipital
> lobe. When you are going after this type of changes, it doesn't make sense
> to use the smallest granularity (pixel). If the change is only 5%, you can't
> find it by pixel-based methods because each pixel is too noisy. TBSS may
> also have too much granularity because each skeleton point may average only
> few adjacent pixels. In this case, an ROI drawn in the frontal lobe and
> occiptal lobe averaging thousands of pixels may be a better method.
>
> On the other hand if a lesion is sharply localized, then you want higher
> granularity.  Pixels averaging always have a danger that abnormal pixels are
> averaged with normal pixels and thus dilute the effect.
>
> Raw pixel-based data is therefore always nice to have. It has low
> sensitivity and you can hardly get any statistical level once you do
> multiple-pixel comparison. But if you lower the statistical thresholding,
> you start to see some tendency. The filtering (pixel averaging) by TBSS may
> help to bring up such information above the surface of statistical power.
> However, you can also argue that if the abnormality is localized to one side
> of the tract, the projection to the center of the white matter core would
> decrease the sensitivity.
>
> Also, you can also argue that the skeletanization is somewhat a blackbox
> operation. When you are looking at the white matter away from the core where
> there are cross-subject variability, I'm not sure how the white matter is
> skeletanized and how they are aligned among subjects.
>
> Again, VBA data is one step closer to raw data before the high order image
> analysis such as TBSS do something on the pixel-by-pixel data. So you can
> argue that VBA and TBSS is not A or B, it is more like A or AxB
>
> Hope it helps.
>
>
> On Fri, May 14, 2010 at 1:41 PM, lion gao <gaolion at gmail.com> wrote:
>
>> Dear Susumu Mori,
>>
>> Thanks very much for your answer, it's very informative!
>> One question is that when you mention granularity and anatomy, it finally
>> all leads to the advantages of TBSS, the skeleton. so my answer looks a bit
>> repeated for a same point. Another one is I do need to find some flaw in
>> TBSS for defence use :P, so I find some limitation from the original paper.
>> BTW, I am NOT familiar with TBSS.
>> http://sirl.stanford.edu/~bob/pdf/DTI/Methods/Smith_TBSS_NeuroImage06.pdf
>>
>> Finally, I arranged some defences based on your comments below. Have a
>> look if interested , or simply skip it :)
>>
>> Thanks again & Best Wishes,
>> Gao
>>
>>
>> It has been quite a time since VBA was popularly used in DTI data, while
>> track-based spatial statistical gains popularity rapid in recent years. To
>> compare the two methods, there two points of views need to the clarified:
>>
>> 1.       Granularity. It represents the extent to which a system is
>> divided into smaller parts. In VBA, the brain is broken into nearly one
>> million of voxels, then it is normalized before comparing with one another.
>> In MRIstudio, it can superimpose the parcellation map and divided the brain
>> into 150 areas, thus it has less granularity than VBA (1 million vs. 150).
>> Whereas TBSS re-registers nearby voxels to the skeleton of white matter,
>> reduces the granularity and increases the accuracy of registration.
>>
>> 2.       Anatomy. In VBA analysis, computer algorithm registers the
>> voxels automatically, without considering any anatomical information.
>> Another new method of Tissue-specific, smoothing-compensated (T-SPOON) can
>> improve the tissue specificity in VBA method and compensation for images
>> smoothing. TBSS can also apply some anatomical information, i.e., the
>> skeleton of white matter, thus the registration would be better than totally
>> automatic VBA method.
>>
>>
>>
>> TBSS definitely has its advantages in DTI analysis, nonetheless, it is not
>> without flaw or limitation: 1.Partial volume effect still exists in TBSS
>> method, and the problem may be greatly exacerbated in spatial smoothing; 2.
>> Increased head motion can increase image blurring and bias FA value; 3. In
>> regions of crossing tracts or junctions, TBSS may misinterpret the change of
>> tracks in junction as apparently reduced FA; 4. Finally, in patients with
>> apparent pathological changes, TBSS may exclude the areas from analysis.
>>
>>
>>
>> There are always advantages and disadvantages of each method in DTI
>> studies. TBSS as new method, may be generally more accurate than VBA method,
>> although it needs more sophisticated data projection approach in the future.
>> We also need to keep in mind that DTI is based on information of water
>> diffusion, it may help to screen and generate hypothesis rather than to draw
>> a final conclusion.
>>
>>
>>
>>
>>
>>
>>
>>
>> -------------------------------------------------------------------------------
>>
>>
>>
>> On 13 May 2010 23:44, susumu mori <susumu at mri.jhu.edu> wrote:
>>
>>> Good question Gao.
>>>
>>> Here is my thought (my personal opinion, of course);
>>>
>>> 1) you are comparing VBA and TBSS. These are methods to "define
>>> corresponding pixels (or areas) across subjects", so that you can compare
>>> pixel numbers such as FA and MD among different brains. This is called
>>> "registration".
>>> 2) These methods can be classified from different point of views;
>>> 2-1) granularity: One extreme is to define the entire brain as one ROI.
>>> You can get the whole brain volume, whole brain FA, or whole brain
>>> histogram. While there is not much use of this approach, it is precise (do
>>> 10 times and you get the same results) and accurate (nobody makes mistakes
>>> about where is the whole brain except for some ambiguity about where you cut
>>> the ROI in the brainstem). The other extreme is the pixel, which is the
>>> smallest unit. This mean, you identify the corresponding pixels across
>>> subjects. Once you map the entire 1 million pixels in one brain to the
>>> other, it is the same as transforming one brain to the other (two brains now
>>> have the same shape). This approach is called "normalization" and, of
>>> course, not accurate because it is not possible to completely solve the
>>> system and accurately map all 1 million pixels. VBA and TBSS are based on
>>> this normalization procedure. There are methods to ameliorate this accuracy
>>> issue. Usually VBA uses filters to blur the information. In my
>>> understanding, TBSS "re-register" nearby pixel information to the core of
>>> the white matter, which could be considered as a sort of filtering, reducing
>>> the granularity and hopefully increasing the accuracy.
>>> 2-2) Anatomy: When we do normalization, computer algorithm do not care
>>> about anatomy. It just does whatever it thinks best to register pixels. This
>>> is the pixel-based analysis. On the other hand, manual ROI is usually based
>>> on anatomical information we can perceive. This is anatomy-based analysis.
>>> Tractography-based analysis can also be considered as a kind of registration
>>> method. We do, for example, tractography of the cortico-spinal tract in 10
>>> subjects. Then we can define a group of pixels that belong to the CST and
>>> compare the pixel values. In this way, we define a specific area across
>>> subjects based on anatomy.
>>> 3) In MriStudio, DiffeoMap does pixel-based registration just as VBA. You
>>> can do VBA analysis. In addition, you can superimpose our parcellation map
>>> and divide the brain into about 150 areas. In terms of granularity, it is
>>> much less than VBA (more than 1 million pixels vs 150 areas). This is also a
>>> conversion to pixel-based to anatomy-based analysis.
>>> 4) Now going back to your question, VBA and TBSS are looking at the same
>>> data with different point of view. First of all, the granularity is
>>> different; TBSS reduces the information to the white matter core. Also, VBA
>>> is completely pixel-based but TBSS, which is not completely anatomy-based
>>> but has some anatomy-based factors by reducing the information to the core
>>> of the white matter.
>>> 5) In my opinion, all methods described above have advantages and
>>> disadvantages. I don't think any one of them is better than the other.
>>> Quantification based on location information is definitely one of the most
>>> difficult problems we are all facing.
>>> 6) On the other hand, your reviewer is correct, in a sense that it is
>>> always important to compare different results to enrich your data and
>>> interpretation. This is especially true if there is a tool widely used like
>>> TBSS. However, it is not like, one method should be treated as the gold
>>> standard and other approaches should give a similar result. If you compare
>>> VBA and TBSS, you likely to get different results because as explained above
>>> they are operating at the different granularity, precision, and accuracy.
>>> MRI image anlaysis is very often a screening and hypothesis generating tool
>>> rather than a tool to draw a conclusion. We are simply looking at 6MB
>>> information based on water signal.
>>>
>>> So, in conclusion,
>>>
>>> > No, VBA is not out of date
>>> > Yes, it is a good idea to compare results from widely used tools, but
>>> any normalization-based method should not be considered as a gold standard
>>> in my opinion. We just have to understand how they operate and what are
>>> their advantages and disadvantages.
>>>
>>>   On Thu, May 13, 2010 at 8:07 AM, lion gao <gaolion at gmail.com> wrote:
>>>
>>>>   Dear Experts,
>>>>
>>>>
>>>>
>>>> I have one part of my thesis on DTI data analysis. The method I used is
>>>> the voxel-based analysis (VBA) and I tried MRIstudio as well. One of the
>>>> examiners pointed out that track-based spatial statistics (TBSS), as a state
>>>> of art way, should have been considered.
>>>>
>>>>
>>>> I am familiar with TBSS, only know that it may reduce systemic
>>>> mis-registration inVBA and increasing papers published with the method. I am
>>>> not sure whether:
>>>> 1. VBA has become “out of date”,
>>>> 2. and TBSS has become a “golden standard” for DTI data analysis.
>>>>
>>>> Can someone help to justify the situation or defense a little bit? Thank
>>>> you very much in advance!
>>>>
>>>>
>>>> Best wishes,
>>>>
>>>> Gao
>>>>
>>>> _______________________________________________
>>>> Mristudio-users mailing list
>>>> Mristudio-users at mristudio.org
>>>> http://lists.mristudio.org/mailman/listinfo/
>>>> Unsubscribe, send a blank email to:
>>>> Mristudio-users-unsubscribe at mristudio.org
>>>>
>>>>
>>>
>>> _______________________________________________
>>> Mristudio-users mailing list
>>> Mristudio-users at mristudio.org
>>> http://lists.mristudio.org/mailman/listinfo/
>>> Unsubscribe, send a blank email to:
>>> Mristudio-users-unsubscribe at mristudio.org
>>>
>>>
>>
>> _______________________________________________
>> Mristudio-users mailing list
>> Mristudio-users at mristudio.org
>> http://lists.mristudio.org/mailman/listinfo/
>> Unsubscribe, send a blank email to:
>> Mristudio-users-unsubscribe at mristudio.org
>>
>>
>
> _______________________________________________
> Mristudio-users mailing list
> Mristudio-users at mristudio.org
> http://lists.mristudio.org/mailman/listinfo/
> Unsubscribe, send a blank email to:
> Mristudio-users-unsubscribe at mristudio.org
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.mristudio.org/pipermail/mristudio-users/attachments/20100516/a8f3f3c8/attachment-0001.html 