connectometry parameters & results

[a a]

Hi Frank,

I have tried several parameters in my DTI connectometry analysis using DSI Studio, an have a few relevant questions as follows,

1. How to identify significant tracks? And how to remove non-significant tracks? If the report.txt file says that the connectivity analysis identified tracts A, B, and C with decreased connectivity with a certain cognitive variable (FDR < 0.05), does it mean that these three tracks (A, B, C) are all significant? Some tracts look very thin with only one or two streamlines (counts), and I am skeptical whether they are significant or not. 

2. If a track is significant at permutation count = 5,000, but becomes non-significant at permutation count = 10,000, then which result is more reliable or correct? How to understand this results difference due to different permutation counts?

3. If a track is significant at the t threshold of 4, but not significant at a lower threshold (t = 3), how to understand and interpret the results? 

Any comments are greatly appreciated!

Best wishes,

Jie

[Fang-Cheng Yeh]

Hi Jie,

1. How to identify significant tracks? And how to remove non-significant tracks?

The FDR value indicates whether the tracks are significant ( < 0.05) or not. The non-significant tracks are filtered using a track length threshold.

 

If the report.txt file says that the connectivity analysis identified tracts A, B, and C with decreased connectivity with a certain cognitive variable (FDR < 0.05), does it mean that these three tracks (A, B, C) are all significant? Some tracts look very thin with only one or two streamlines (counts), and I am skeptical whether they are significant or not. 

If the tracks look very thin, I would suggest increase the permutation count or decrease the T-threshold so that more regions can be included with the findings (with the sacrifies of FDR).

 

2. If a track is significant at permutation count = 5,000, but becomes non-significant at permutation count = 10,000, then which result is more reliable or correct? How to understand this results difference due to different permutation counts?

10,000 would be more reliable only if the subject number is large enough to make the statistics converge. If you do not have enough number of subjects, then permutation may fail.

 

3. If a track is significant at the t threshold of 4, but not significant at a lower threshold (t = 3), how to understand and interpret the results? 

It is likely that the finding is very localized and can only stands out at a higher T-threshold, or the data/subject variance is high across brain regions, leading to several region occationally reaching high T during random permutation. This may happen if the subjects are highly diverse, or the subject pool is not large enough. 

Best regards,

Frank

 

Any comments are greatly appreciated!

Best wishes,

Jie

--
You received this message because you are subscribed to the Google Groups "DSI Studio" group.
To unsubscribe from this group and stop receiving emails from it, send an email to dsi-studio+unsubscribe@googlegroups.com.
To post to this group, send email to dsi-s...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[jzhua...@gmail.com]

Hi Frank,

Thank you very much indeed for your comments! When I correlated a behavioral measure with DTI maps across all participants, I found from the report.txt file that "The connectometry analysis identified no track with increased connectivity related to behavioral_measure (FDR=1) and external capsule l with decreased connectivity related to behavioral_measure (FDR=0)".  Does it mean the "external capsule l" is negatively significantly correlated with the behavioral measure? When I looked into corresponding track files, I found only one track or streamline (external capsule l) in the lesser.trk.gz file, but 14 tracks/streamlines in the greater.trk.gz file. It seems that 14 tracks in the positive correlation are more likely to be significant, compared to 1 track in the negative correlation. Could you be kind to give me some hints to understand the results? Many thanks!

Best wishes,

Jie

Hi Jie,

To unsubscribe from this group and stop receiving emails from it, send an email to dsi-studio+...@googlegroups.com.

[Fang-Cheng Yeh]

It seems that the permutation count and seed/track density were not enough. You may increase them in the advanced option to see if it give more results.

To unsubscribe from this group and stop receiving emails from it, send an email to dsi-studio+unsubscribe@googlegroups.com.

[jzhua...@gmail.com]

I used the permutation count of 10,000 and the seed density of 50 which are already the maximal values in the software. Could you please help me increase the limits in your software? Thank you very much indeed! 

[Fang-Cheng Yeh]

Usually we don't need to use such a high permutation count. It seems to me that there could be other problem. Is it possible to send me the connectometry dB and demographic file to check out the problem?

To unsubscribe from this group and stop receiving emails from it, send an email to dsi-studio+unsubscribe@googlegroups.com.

[jzhua...@gmail.com]

Do you mean the xxx.sdf.db.fib.gz file? Mine is very big, around 1.6GB. Could you please suggest a way to transfer this file to you? Thank you very much indeed!

[Fang-Cheng Yeh]

I received the file. I would suggest checking "normalized SDF" and use
a T threshold of 2. The lower T threshold is because the regression
variables have non-linear effect (e.g. age and education), and it is
unlikely you will have a high T-score.

There is a recent update to the connectometry, and you may also need
to update DSI Studio before running the connectometry.

Best regards,
Frank

[Jie]

Thank you very much indeed, Frank! I have tried to use "normalized SDF", and the results look normal now. I tried both T=2 and T=3, and found significant tracts in both analyses. Of course, the tracts generated at T=3 are much fewer than those generated at T=2. Do you mean that age and education are less likely to be normally-distributed so that their effects are non-linear? 

I am sending you my latest analysis results. According to the report.txt file, corpus callosum and cortico thalamic r are significant tracts in the positive correlational analysis. The db.fib.gz file shows a bundle of tracts in the splenium and only one track in the genu, so does the significant effect in "corpus collosum"  actually mean the splenium tracts? Or, does it also include the single tract in the genu? Does "cortico thalamic r" mean right corticospinal tract? There seem to be more corticospinal tracts in the left hemisphere than in the right, but only the right hemisphere tracts are significant. This seems a bit weird. Is there a chance to produce a result table similar to that in SPM by including each tract or tract bundle with its specific details (e.g. size, significance, t or Z value, name)? Do you have any recommendation on software and templates to use in order to identify each tract?    

Another question is the meaning of T threshold and FDR correction. Can they be understood by analogy with canonical BOLD activation analysis that T threshold is similar to voxel-level uncorrected and FDR correction is similar to cluster-level correction? Is T=2 too low to report in formal publications? Do you recommend T=2.3 (which is equivalent to p=0.01), T=3.09 (p=0.001), or other thresholds? Sorry to bother you with so many questions! Any help is sincerely appreciated!

[Fang-Cheng Yeh]

> Thank you very much indeed, Frank! I have tried to use "normalized SDF", and
> the results look normal now. I tried both T=2 and T=3, and found significant
> tracts in both analyses. Of course, the tracts generated at T=3 are much
> fewer than those generated at T=2. Do you mean that age and education are
> less likely to be normally-distributed so that their effects are non-linear?
>

The linearity is not really related to whether the variables are
normally distributed.

A linear relation here means that age=20 and age=30 will have a ratio
of 1.5 in terms of their connectivity difference. This is unlikely to
be true.

> I am sending you my latest analysis results. According to the report.txt
> file, corpus callosum and cortico thalamic r are significant tracts in the
> positive correlational analysis. The db.fib.gz file shows a bundle of tracts
> in the splenium and only one track in the genu, so does the significant
> effect in "corpus collosum" actually mean the splenium tracts? Or, does it
> also include the single tract in the genu? Does "cortico thalamic r" mean
> right corticospinal tract?

The track recognition provides by DSI Studio is a rough suggestion.
What you can do is converting finding to a seed region using
[Tracks][Tracks to ROI] to track the entire pathway.

There seem to be more corticospinal tracts in the
> left hemisphere than in the right, but only the right hemisphere tracts are
> significant. This seems a bit weird.

> Is there a chance to produce a result
> table similar to that in SPM by including each tract or tract bundle with
> its specific details (e.g. size, significance, t or Z value, name)? Do you
> have any recommendation on software and templates to use in order to
> identify each tract?

The length of a track will tell its significance value (FDR) here.
Usually, tracks with a longer length will have a lower FDR value. You
may apply a length threshold using [Tracts][Delete short tracks] to
tracts with a longer length and then look up their FDR values in the FDR
table.

>
> Another question is the meaning of T threshold and FDR correction. Can they
> be understood by analogy with canonical BOLD activation analysis that T
> threshold is similar to voxel-level uncorrected and FDR correction is
> similar to cluster-level correction?

Yes, their share the same analysis paradigm.

> Is T=2 too low to report in formal
> publications? Do you recommend T=2.3 (which is equivalent to p=0.01), T=3.09
> (p=0.001), or other thresholds? Sorry to bother you with so many questions!
> Any help is sincerely appreciated!
>

I think T=2 is okay because of the nonlinearity nature of your
variables (age, education).

Best regards,
Frank

[Jie]

Thank you very much indeed, Frank!

If some tracks are reported in the report.txt file as significant (FDR < 0.05), does it mean this effect is based on all tracks in the greater.fib.gz (or lesser.fib.gz) file as a whole, although the names of some minor tracks with very few streamlines are not mentioned in the report.txt file? Or, is this significant effect driven only by the major tracks (while the minor tracks are not significant)?  

On your website, you mentioned that "For studies aiming at findings which tracks are correlated with the study variable, use "FDR" as the threshold. DSI Studio will capture tracks with FDR lower than the predefined threshold value (e.g. 0.05)". Where is this FDR option in the DSI Studio GUI? Many thanks for your great help!

[Fang-Cheng Yeh]

On Fri, Apr 13, 2018 at 7:06 PM, Jie <jzhua...@gmail.com> wrote:
> Thank you very much indeed, Frank!

> If some tracks are reported in the report.txt file as significant (FDR <
> 0.05), does it mean this effect is based on all tracks in the greater.fib.gz
> (or lesser.fib.gz) file as a whole, although the names of some minor tracks
> with very few streamlines are not mentioned in the report.txt file?

Yes. the FDR is calculated from all tracks

>
> On your website, you mentioned that "For studies aiming at findings which
> tracks are correlated with the study variable, use "FDR" as the threshold.
> DSI Studio will capture tracks with FDR lower than the predefined threshold
> value (e.g. 0.05)". Where is this FDR option in the DSI Studio GUI? Many
> thanks for your great help!

It is available in the recent release. Please update DSI Studio to see
if the GUI shows options for FDR threshold.

Best regards,
Frank

[Jie]

Thank you very much indeed! I found the FDR option in the new version. 

If track A is significant at a low track density threshold (e.g. 5 seeds per voxel), does it mean that the effect of track A is very strong and stable? In contrast, if track B is only significant at a high track density threshold (e.g. 20 seeds per voxel), but not at a low density threshold, then it means the effect of track B is subtle and not stable, right? Could the effect of track B false positive? Thanks.

[Fang-Cheng Yeh]

> Thank you very much indeed! I found the FDR option in the new version.
> If track A is significant at a low track density threshold (e.g. 5 seeds per
> voxel), does it mean that the effect of track A is very strong and stable?

It seems so, but the "strong and stable" is determined by the FDR.
The lower track density does not necessarily suggest the significance.

> In contrast, if track B is only significant at a high track density
> threshold (e.g. 20 seeds per voxel), but not at a low density threshold,
> then it means the effect of track B is subtle and not stable, right? Could
> the effect of track B false positive? Thanks.

Yes. you may check out the "length" of the track B and look up its FDR
in the connectometry GUI (second tab of the connectometry plot)

A shorter track suggests a higher FDR (more likely to be false positive)

[Jie]

Thank you very much indeed! Your comments are very helpful!

As you mentioned before, some regressors (e.g. age) have linearity problems in the connectometry analysis, so the option of "normalize SDF" is recommended. Are there any gold standards on when "normalize SDF" should be used? For regressors with large variations across participants (e.g. age), "normalize SDF" is needed for sure. However, for other regressors with small variations across participants (e.g. some cognitive variables), I feel puzzled in making a judgment whether it is appropriate to use "normalize SDF" or not. I did some tests, and found that the effects of some cognitive variables remain largely constant no matter "normalize SDF" was used or not. For other cognitive variables, the effects were largely changed or even reversed (from negative to positive, or vice verse) when "normalize SDF" was used, compared to when "normalize SDF" was not used. Should "normalize SDF" be used for a categorical regressor (e.g. a group variable of patients vs. controls)? Could you be kind to provide further suggestions on this issue? Sorry to bother you so much! Any tips are greatly appreciated! 

[Fang-Cheng Yeh]

"Normalize SDF" affects how diffusion signals are calibrated/scaled,
and there is no strict guideline for when it should be on and off
because it may change due to scanner setting and also the diffusion
pattern of the scanning targets.

My experience is the following:
(1) If the direction of the correlation is wrong (this requires your
prior assumption for judgement), switch "normalize SDF" to the other
state (on->off, off->on)
(2) If normalize SDF does not affect the results or FDR, turn it off.
(3) If normalize SDF improves the FDR, leave it on.
(4) Make a consistent choice for the same data set. (i.e. do not turn
it on when correlating with certain variables and off for the others)

Hope this helps.

[Jie]

Your comments are very useful. Thank you very much indeed!