Hi Robert,
> Indeed, there are considerablly differing results between upward (E) and downward (A-C) tracking what matters the connectivity probability of one spinal voxel, but also between A and B I get differing ratios between the different motor areas in one subject. So, the correct approach seems to be crucial and if you don't mind, I'd like to add just a few questions.
>
> A) Is this approach considering seed mask size differences by dividing by (n seed voxels x 5000) enough? Target mask is spinal of the same size for every subject.
You need to divide by the number of samples to get comparable probability values cross subjects (#sample=#voxels X #sampesPerVoxel).
However, even if you do that, your results will still depend on the size of your seed mask, in a non-trivial way. This is not an issue, as long as your seed masks represent the same functional area across subjects. I.e., it is more important to match rois across subjects in terms of their anatomy/function than # of voxels (this is the tricky problem!).
> B) This is what I really don't understand: if I get 3400 hits in a spinal voxel from a 500 voxel cortical seed mask, under certain brain waypoint and exclusion masks a waytotal of 5000 tracts, there is a big difference, so non equivalent (?) between both probabilities
>
> P1 = 3400 / 5000 (only waytotal, would I tremendously overestimate?
>
> P2 = 3400 / 466 x 5000
>
> How does the single devision by waytotal account for exclusion criteria, aberrant tracts and seed size differences? Am I missing an important clue?
There is no under- or over- estimation, you use a different formula for calculating a different probability.
The main difference is between calculating the probability of seeding from A and ending in B (the standard fdt_paths output), and adding priors using various masks.
P2 is the right thing to use if you want to compare values across subjects. It normalises the distribution with respect to the number of samples drawn from that distribution (adding prior mask is similar to rejection sampling).
>
> C) I thought about a "two-way" approach for more robust tracking.
> If this doesn't make a sense, could you please explain what file I have to read after multiple-mask approach and how I have to normalize for different seed mask sizes to determine the probability P of a single spinal voxel to be robustly connected (multiple mask) to M1. I want to generate a connectivity profil for each voxel.
the waytotal text file will have the number of samples drawn from M1->Spine and from Spine->M1.
> Finally, how would you do that kind of analysis??
Exactly as you did:
Be very careful in defining your seeds and targets.
Use prior masks (inclusion/exclusion/stop) to inform the trajectory of interest.
Divide by waytotal :)
Saad.
> Thank you so much for your help!!!
>
> Cheers
> Robert
>
>
>
>
>
> ======================
>
> Hi Robert,
>
> In my opinion:
>
> A is fine.
>
> B is equivalent to dividing by waytotal (condition on that particular tract), which is what we would recommend if you want the probabilities to be more comparable between subjects and less dependent on seed/target mask sizes.
>
> C and D do not make sense to me...
>
> It would be interesting to compare E and B, although in E I would also divide by waytotal to get around the target size issue.
>
> Cheers,
> Saad.
>
>
> On 21 Oct 2010, at 20:38, Robert Schulz wrote:
>
>> Dear FSL experts,
>>
>> having gone through quite a lot of older discussions I am still wondering what way would the best to work out voxel connectivity profiles. I am quite confused since I red that division by waytotal should be enough in unidirectional tracking.
>> So if you don't mind, I'd like to explain my 5 actual approaches for connectivity probability of each spinal voxel to be connected to M1.
>>
>> just for information:
>> downwward tracking = M1 seed mask (500 voxels, 5000 samles each) to spinal mask as classification mask
>> upward tracking = spinal (230) to M1 classification mask
>> waypoit masks: peduncular mask, spinal mask (in upward M1)
>> extending exclusion mask (hemispheres, basal ganglia, cerebellum)
>>
>>
>> Calculation of every spinal voxel connectivity probability is achieved via:
>> -----------------------------------------------------------------------------------------
>>
>> A) (Spinal voxel's downward fdt_path value) / 500 x 5000
>>
>> B) (Spinal voxel's downward fdt_path value) / (only non-zero M1 voxels as read in downward's seed_to_spin x 5000)
>>
>> C) (Spinal voxel's downward fdt_path value) / (only non-zero M1 voxels as read in upward's fdt_path x 5000)
>>
>> D) (Spinal voxel's downward fdt_path value) / (only that M1 voxel that are non-zero as in B and (!) C x 500)
>>
>> The reason for approaches B-D was to exlude those M1 voxel from the normalization that were covered by the mask but are likely not sending and/or receiving paths to/from the spinal mask. To prevent undercorrection of the spinal voxel connectivity profiles.
>>
>> E) Simply upward-tracking, reading upward's seed_to_M1 voxel values / 5000. In this case, do I have to consider different cortical target mask sizes? Does they impact on tfhe spinal voxel connectivity probabilities?
>>
>>
>> I would be very grateful if you could comment shortly on each approach. I know they are arbitrary, but I really want to consider as much as possible differing parameters to get robust probabilities.
>>
>> There are some more ways to calculate robust VCP I could think of (e.g. A x E, or further normalization for seed masks averaged FA), but for now I am confused enough. Can someone help me out here?
>>
>> Looking forward to your comments?
>>
>> Best
>> Robert
>>
>
> --
> Saad Jbabdi
> University of Oxford, FMRIB Centre
>
> JR Hospital, Headington, OX3 9DU, UK
> (+44)1865-222466 (fax 717)
> www.fmrib.ox.ac.uk/~saad
>
>
--
Saad Jbabdi
University of Oxford, FMRIB Centre
JR Hospital, Headington, OX3 9DU, UK
(+44)1865-222466 (fax 717)
www.fmrib.ox.ac.uk/~saad
|