 |
 |
Hi Stefano Before I (attempt to) answer your question, let me first make an important remark: you are saying that your experiment compares deterministic and probabilistic tractography. I would say that more importantly, it is comparing one versus two fibre models of local orientations. Probabilistic simply means that you account for the reproducibility of the tracking trajectories - if you have no noise and your model is perfect, then deterministic and probabilistic will give the same results, as long as they are using the same model. I bet the difference you see between the two are mainly due to modelling crossing fibres, especially if you are looking for parieto-frontal connections; some of those are notoriously difficult with single fibre models. The other point, which relates to your part II email, is that since you are trying to compare the *location* of streamline connections between groups, you want to be as accurate as possible about these locations. In particular, you want to avoid repeated resampling of the tracking results. This is why probtrackx allows you to output your results directly as a histogram in the space that you are interested in (MNI), without resampling the data twice, as you would do if you ran tracking in diffusion space then transformed into MNI. So you should definitely run your analysis directly in MNI space. Other than that, I think your method is fine. I don't see a problem with comparing binary maps as long as you use non-parametric stats (e.g. randomise). I am not sure how cluster thresholding will behave in practice, but it should behave similarly to VBM... Now, about interpretation. You are comparing tract location across two groups, and you want to know what are the driving factors for the differences. I would look for the following clues: - Is it a difference in the location or a difference in the "width" of the tract? Most likely, your statistic will be very sensitive to changes in tract width. Differences are easier to detect at the borders between tract and no-tract. If the centre does actually change, it will be harder to detect I guess - Differences can be due to one group having very reproducible results across subjects, and the other scattered results (e.g. tumour patients). You can evaluate this by eye, or if the differences are subtle, you can do the following tests: * If the tract is a bundle, then you can plot the x,y,z coordinates of the mean and std location as a function of distance from seed. (the --pd option in probtrackx can be very useful for this) - this may also allow you to visualise displaced tracts (as opposed to change in width) * If the tract is not bundled, then you can use prior ROIs to separate it into different bundles before doing the plots suggested above * You can also plot tract orientation (mean+std) as a function of the distance from seed. This may show if the groups differ in fibre spread. This one may be tricky to do with probtrackx as it does not output the tract orientation - I admit.... Maybe you can use the tangent to the mean x,y,z as a guide for picking the appropriate local fibre population? - Be careful to check that some of the differences are not simply due to gross mis-registration to MNI space I hope this helps Saad On 5 Jan 2012, at 18:55, Marenco, Stefano (NIH/NIMH) [E] wrote: > Dear FSL list, I have two groups, normal controls and “patients” and I run probabilistic (bedpost default values) and deterministic (DTI studio) tractography from a seed in the intraparietal sulcus. > > In order to do this, we started from a seed location in MNI space, transformed it into individual space and grew a spherical ROI around it. Then we ran the tractography and binarized the tracts (choosing a reasonable threshold for the probabilistic tractography). Then we warped the tracts back into MNI space and created a mask to limit statistical testing 1) to areas that had a tract present in at least one subject and 2) in superior portions of the brain (based on a prior hypothesis). In this mask we ran randomize with an arbitrary cluster finding threshold of t=3 (due to the binary nature of the images we could not get TFCE to work). > > I get significant differences in the location of the tracts with both modalities, but in different regions. The deterministic tractography differs very close to the seed in the corona radiata where multiple fiber bundles mix. The region of significant difference in the probabilistic tractography is more anterior in the superior thalamic radiation. > > I found these results interesting b/c I thought that they might be interpreted as complementary findings indicating differences due to the orientation of the primary eigenvector for the deterministic tractography (overcome by the presence of crossing fibers that are then detected and tracked through by probabilistic tractography) and in more complex differences along the tract for probabilistic tractography, possibly related to the cumulative effects of tracking along primary and secondary fibers. > > Does this interpretation make sense? To verify this, I thought that for the deterministic difference I should find different orientations of the primary eigenvector at or around the area of difference in tractography, but how can I determine what causes the difference in probabilistic tractography? > > Thanks for your help in advance, I am going to send a part II where we started looking at one possible source of error in trying to understand these results. Stefano Marenco -- Saad Jbabdi University of Oxford, FMRIB Centre JR Hospital, Headington, OX3 9DU, UK (+44)1865-222466 (fax 717) www.fmrib.ox.ac.uk/~saad