Question to the instructors on the project

[hveravallaskreppur]

Hi @matthew-brett, our project is currently rather basic. It runs but only prints out outliers in each scan according to x nr of std from the mean. We intend to prove it using other outlier detection methods (like IQR) and provide a guesstimate on why we think it is better.

A few questions on the project: i) our understanding is that the project is focused our pre-processing of fMRI data. Meaning that if an fMRI project were split into three phases: 1) data collection (scans), 2) pre-processing and 3) statistical analysis (HRF etc.), this project is purely focused on the middle part, i.e. pre-processing – therefore we shouldn't be applying SPM or GLM or anything to the scans, right? ii) we didn't have time to dive deep into transformations in the course (last lecture) are you expecting us to to do any movement corrections before applying outlier detections? iii) any hints or guidance on what is a "pass" and what is a "well done" in terms of project delivery would be much appreciated :)

[hveravallaskreppur]

Hi @matthew-brett, I sent the above question on Wednesday, did you see it? Adding @oesteban @effigies

I don't think we'll be able to deliver the project today.

[matthew-brett]

Sorry for the delay - I'm just getting ready to start my teaching this term, and it's been busy.

A few questions on the project: i) our understanding is that the project is focused our pre-processing of fMRI data. Meaning that if an fMRI project were split into three phases: 1) data collection (scans), 2) pre-processing and 3) statistical analysis (HRF etc.), this project is purely focused on the middle part, i.e. pre-processing – therefore we shouldn't be applying SPM or GLM or anything to the scans, right?

That's your call - I said something about this in the sessions - but you can do any level, from just considering the scans as they are, without further processing, all the way through to testing your techniques on the error data from the statistical model - that is, seeing whether the scan is an outlier, taking the statistical model into account.

ii) we didn't have time to dive deep into transformations in the course (last lecture) are you expecting us to to do any movement corrections before applying outlier detections?

We didn't get into that, so no, I'm not expecting you to do that - of course, you might consider doing that, if you know how to using other tools, such as SPM or FSL, and you have time to investigate.

iii) any hints or guidance on what is a "pass" and what is a "well done" in terms of project delivery would be much appreciated :)

Pass is spending some time thinking about how to calculate measure to detect outliers, and some evidence that you've investigated different methods and evaluated them for performance on some metric you think is sensible - even if this is reproducing what looks to you - by eye - to be outliers.

Well done is going deeper into the problem, perhaps by considering pre-processing, perhaps by taking the statisical model into account, perhaps by comparing to other off-the-shelf methods - and comparing performance in a more formal way between methods - your methods, or other methods.

[hveravallaskreppur]

That's very helpful, thank you well plough on, Best David

---

From: Matthew Brett @.> Sent: Friday, September 29, 2023 11:38 AM To: nipraxis-summer-2023/diagnostics-bold_but_better @.> Cc: D. Gudjonsson @.>; Author @.> Subject: Re: [nipraxis-summer-2023/diagnostics-bold_but_better] Question to the instructors on the project (Issue #27)

Sorry for the delay - I'm just getting ready to start my teaching this term, and it's been busy.

A few questions on the project: i) our understanding is that the project is focused our pre-processing of fMRI data. Meaning that if an fMRI project were split into three phases: 1) data collection (scans), 2) pre-processing and 3) statistical analysis (HRF etc.), this project is purely focused on the middle part, i.e. pre-processing – therefore we shouldn't be applying SPM or GLM or anything to the scans, right?

That's your call - I said something about this in the sessions - but you can do any level, from just considering the scans as they are, without further processing, all the way through to testing your techniques on the error data from the statistical model - that is, seeing whether the scan is an outlier, taking the statistical model into account.

ii) we didn't have time to dive deep into transformations in the course (last lecture) are you expecting us to to do any movement corrections before applying outlier detections?

We didn't get into that, so no, I'm not expecting you to do that - of course, you might consider doing that, if you know how to using other tools, such as SPM or FSL, and you have time to investigate.

iii) any hints or guidance on what is a "pass" and what is a "well done" in terms of project delivery would be much appreciated :)

Pass is spending some time thinking about how to calculate measure to detect outliers, and some evidence that you've investigated different methods and evaluated them for performance on some metric you think is sensible - even if this is reproducing what looks to you - by eye - to be outliers.

Well done is going deeper into the problem, perhaps by considering pre-processing, perhaps by taking the statisical model into account, perhaps by comparing to other off-the-shelf methods - and comparing performance in a more formal way between methods - your methods, or other methods.

— Reply to this email directly, view it on GitHubhttps://github.com/nipraxis-summer-2023/diagnostics-bold_but_better/issues/27#issuecomment-1740750367, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AH7S4LHXAEQB67ALKKSBNJTX42XNFANCNFSM6AAAAAA5JE6E7I. You are receiving this because you authored the thread.Message ID: @.***>

[hveravallaskreppur]

@matthew-brett I am struggling to understand the events files with the scans. The TR is clearly not 2.5 is in the course examples. Looking at the events file for subject 1:

onset duration trial_type 0 1 2 3 0.000 0.350 Words mist 1 NR NR 1.000 0.350 Words otter 1 NR NR 2.000 0.350 Words otter 1 34 577 3.000 0.350 Words verse 1 NR NR 4.000 0.350 Words pail 1 NR NR ...

I'm trying to work out a convolved time course just looking at the first two columns, onset and duration. But how could this be? Is the TR =0.5sec or something and the participant bombarded with stimulus (events) every second??

[effigies]

@hveravallaskreppur The task stimulus can be presented at a different rate than the scanner is recording. Generally, you will need to construct a stimulus time course at a fairly high temporal resolution (perhaps 0.01s), convolve it with an HRF, and then downsample the convolved timecourse to the sampling rate of the BOLD image.

[matthew-brett]

Following on from Chris' response - the stimuli were words presented every second. You can actually find the correct TR in the Nibabel header data (pixdim[4]) - it's 3 seconds. You might want to look at https://textbook.nipraxis.org/non_tr_onsets.html

[hveravallaskreppur]

Thanks for a quick reply, @effigies, @matthew-brett . I think I understand, will make some attempts. Trying to use GLM as the the outlier method metric and compare F statistic before and after removing outliers, per method. Hope that makes sense

[hveravallaskreppur]

@matthew-brett can I assume that the fourth column in the .tsv is the amplitude? I can see that some rows have "NR" (not recorded?) ?

[matthew-brett]

The columns in the file have names:

onset   duration        trial_type      0       1       2       3

I happen to know that the "trial_type" column is wrong, you should ignore it. Column 1 gives the trial type where 1 = Words, 2 = Consonant strings, 3 = objects, and 4 = scrambled objects. Column 0 gives the name of the stimulus presented. Column 2 gives the button pressed or NR if no button pressed (they always seem to press button 34). Column 3 gives the reaction time in milliseconds or NR if no button pressed.

[hveravallaskreppur]

answers provided with thanks