 So, I hope that we all know that fMRI is a relatively poor temporal property, so as most people know, fMRI has a poor temporal resolution and has poor temporal accuracy in the detected signal. So, in part, for sure this has to do with the vascular nature of the fMRI signal, so the bold signal is measured as a vascular signal and part of this temporal nexity is due to this vascular nature. However, what I'm going to try to convince you of here is that part of the temporal in accuracy is also due to particular methodological decisions that we take in the fMRI data analysis pathway. And so, improving the temporal accuracy is important for event-related fMRI for studies of functional connectivity and for studies that look at the coupling between neuronal activity and vascular responses. All these three techniques are interested in the precise dynamics of this fMRI signal and so, improving the temporal accuracy by which we can extract this signal from the data would be relevant to all these types of different studies. So, here I'm trying to convince you that there's a new method for doing this. So, basically, this talk has two parts. First, I'll just talk a little bit about the standard way of creating these whole-brain volumes and the standard way and the new way to do that. And then in the second part, I'll do a little bit of a try to evaluate this new method in the context of simulated data and some real-world data. So, a key part of fMRI data acquisition is that the fMRI data is collected in terms of slices. So, at each point in time, when an fMRI experiment is going on, the machine is sampling just a little bit, a little tiny part of the brain. And so, when you're in the machine, it is sampling sequentially or in some other form, but it's sampling at each point in time just a little part of the brain. And so, for example, if you have just a brain with just three slices, you would first sample this part, then this part, then this part, and then the bottom part again, that's basically how fMRI data is acquired. This is just a different way to represent that. Here you have time, you have the three slices, and you can see that at each point in time, you just have a little portion of the brain as your data. So, in this raw format of your data, it's actually impossible to do whole-brain analysis, because at no point in time, you have a whole brain available. So you need to do something. The standard solution to do this, to be able to do whole-brain analysis, is to basically shift these slices in time. So the idea here is to just relabel the time point at which these samples were taken. So you would say that actually, slice number one is not acquired at time point zero, but it's actually acquired at time point one. And so what happens when you do that is that you create whole-brain volumes. So you create brains in which all the slices are assumed to be acquired at some point in time. So you can see that you have a whole-brain volume. So you can see that you have a whole-brain volume. In which all the slices are assumed to be acquired at some point in time. And so I should say that this step is often confusing, but this is a step that happens basically when you convert your diagrams to your NIFTs. And this is the step D, this panel D, is when you actually start your fMRI analysis. It has nothing to do with preprocessing. This is really the first thing that you do before you start your analysis. And so there are two major problems with this solution. And that is that these volumes that we have in this way, they contain temporal distortions. So it is not true that this slice here was acquired at time point one. And that this slice here was acquired at time point one. Because you have just relabeled them. And so it's temporally not accurate this data. The second problem is that you lose temporal resolution. And that you actually sample data at time point zero and at time point two and at time point three. But unfortunately in this representation of the data, that temporal resolution is lost. And this leads to a typical resolution in fMRI that is only order of seconds. So typical fMRI experiments that have a TR that is around between two and three seconds. And that have a number of 30 slices or more. You have a temporal resolution that is around on the order of seconds. So what we have come up with is to create whole brain volumes that are composed out of slices that are all acquired at the same point in time but relative to a stimulus. So this method requires that in your fMRI experiment you should present stimuli. And then you compose your volumes based on slices that were acquired at the same point in time relative to a stimulus. So to explain that a little bit. Imagine here you have again this fMRI data but now you have three stimuli. So S1, S2 and S3 are three stimuli. And they're presented exactly in phase with the acquisition of these three slices. So this means that stimulus one is presented exactly when you are sampling the data from slice number one. So this means that slice number one, that data is sampled exactly zero milliseconds after the presentation of the stimulus. So simultaneously with the presentation of the stimulus. For stimulus number two the second slice is sampled exactly zero milliseconds after the presentation of a stimulus. And likewise here slice number three is sampled exactly zero milliseconds after the presentation of a stimulus. So what you then can do is combine these three stimuli these three slices that were acquired for these three stimuli and Create a whole brain volume that is composed out of slices that were all acquired at exactly the same point in time relative to stimulus and actually if you fill out the data More because you were presenting these stimuli in face with the slices You actually get a whole brain volume at each time point where you sample data So you get a much you get more accurate data and at a higher temporal resolution So this method solves these two aforementioned problems that you With this method you don't no longer have these within volume temporal distortions Volumes are composed out of slices that are all acquired at the same point in time And you have a much higher temporal resolution. So with normal fMRI data when you have about 30 slices and the TR of two seconds you can get Up to a maximum temporal resolution about 70 milliseconds Okay, so I hope that is clear That's basically the the the theory part So this method really fundamentally differs from standard fMRI methods It differs in how volumes are created. It also differs in the way the statistical extraction of the signal works. So I Don't have really time to go into this But I Just want to say that when you do normal fMRI analysis, you have a time series you model an entire imaging run In your model in this case, we have Epochs so we have a data that looks much more like EEG data Where you have a certain a certain epoch that is fixated all the stimulus and you You model each time point in this epoch separately You can ask me about this more in detail if you want so here I just wanted to Evaluate this method so and we did this in the context of Simulated data, so of course the advantage of simulated data is that you know what the signal is that you want to detect and you can use Your method this slice base method that we have Developed and compare it to how well it does against the standard method. So the standard method of extracting fMRI signal Is this method called fur or fear? and Yeah, this is just the way that The standard ways of extracting bold signals So we are using that as our Stand as our comparison The idea behind the simulation is to Create a large patch of fictitious neural tissue and we assume that across this entire fictitious neural tissue There is the same hemodynamic response so We would expect that if fMRI were to sample this This this fake this neural tissue you would exact we would expect to see the exact same signal coming from each part of this neural tissue because The hemodynamic response in the signal is exactly the same everywhere Okay, but so we are sampling this In this particular simulation we just use three slices We have a TR three seconds and we just do sequential sampling So slice number one slice number two slice number three To keep things simple So we using this idea we generate basically an fMRI experiment We use 60 stimuli as a slow event related type experiment where you have long pauses between each Simulus presentation to let the hemodynamic response go back to baseline Each stimulus generates exactly the same hemodynamic response using this double gamma HRF and so again each Sample each slice samples exactly the same hemodynamic response function So the questions we had here was how well does each method the slice-based method or this fear method? We cover the ground truth signal And we can look at that by a simple correlation between our ground truth signal and our observed bolt signal We also looked at how well is the signal similar across the three slices? So each slice Samples exactly the same hemodynamic response And so you would expect to see the exact same signal across these three slices So this is results from this fear method So you see the in color you see the results from the three slices. So this is the bolt signal is extracted from this Fictitious neural tissue with this fMRI parameters that we that we that we chose So the red is the first slice green the second blue is the third slice and a dash line is the ground truth signal And so you can see that actually there's this temporal shift Of the of the of the signal across these three slices and this actually makes sense because as I said before The standard method of volume creation relies on creating volumes in which you in which you not exactly Which the slices are not acquired at the same point in time So you so you get here these temporal distortions within each volume and that's what you recover here That's what you see By contrast in the slice-bath method here you see that actually the signal is captured extremely accurate So there's a 99% accuracy And by which the signal is extracted and that is because each volume here is composed out of exact exactly Slice that were slices that were acquired at exactly the same point in time relative to the stimulus So you have no temporal distortions also the The similarity of the signal across the three slices here is low and it's much higher in the slice-based method You can increase the temporal resolution of the data by increasing the number of slices. This is a common method And by jittering the stimulus So you don't present the stimulus exactly at the TR, but you jitter this a little bit This is a common method to increase temporal resolution Again, even with 250 milliseconds you can see that the slice-based method retains about 99% accuracy compared to the standard method where really the performance goes down and this is about 77% Correct So in this particular simulation the slice-based method really more accurately extracts this bold signal with Near perfect accuracy. It's also the accuracy across the slices is much higher If you look at all the simulations that I did so I did much more simulations than the one I just presented here You see about a 25% improvement In addition the variability that the signal is detected is much smaller So it's also a more precise method In terms of real-world data We did We took data from a picture naming task So this is a very simple task with subjects are in the scanner They see a picture and they just have to say what it is So for example in this case you see a picture of a horse and you see a horse So this is also a slow event related design with long pauses in between To really let the bold signal come back The Logic here was that previous research has shown that in this task in the picture naming tasks There's a large portion of motor cortex that is active You're speaking your mouth is moving so motor cortex is supposed to be highly involved in this and we looked at We looked at three slices that were Centered on motor cortex assuming that each slice would see a similar hemodynamic response Okay, so that was the key assumption that we had three slices and that this whole patch of new motor cortex here would Show a similar a hemodynamic response FMI parameters the only thing that is important here is that this is a really standard EPI protocol So this is not this method is applicable to very standard Run-of-the-mill EPI sequences. You don't require multiband or anything like that. We didn't do a Much pre-processing we didn't use any smoothing We just extracted the signal from three slices in one single voxel voxel with the highest signal intensity basically So what you can see here? This is from a single subject This is the standard method and you can see I hope that you can see that the single actually is a is more Diverse across the three slices and here you have a more coherent version of that signal so Assuming that the same hemodynamic response underlies these three slices that are very close together here in motor cortex You would expect a more similar bold signal across these three extracted slices If you do that across the entire group of subjects, so 30 participants There is a significant improvement and enter slice correlation for the slice-based method compared to the standard method also the Slice-based method detects less peaks So there's a different peak in one slice versus another peak You would expect all slices to show the same peak if the same signal was underlying each slice Okay, so this method allows to extract the whole brain bold signal at With high accuracy and with high tempo resolution, so I think this is important for these three methods so event-related fMRI functional connectivity and your vascular coupling I think this has applications in both in also in clinical neuroscience where Accurate action of these bold signals could be used as early biomarkers However, I should make a big Beware here This is to say that we can now extract high tempo resolution bold signals, but we should not equate The bold signal with the time course of neuroactivity these things are not the same and there's a lot of studies that have looked at how well Bold signals may map on to a neural activity and if you're going to use this method You should really be aware of the limitations of using that So to conclude this is a method that is really fundamentally different from a normal Method of analyzing fMRI data the slices and by which you can you can create whole brain volumes That is really different the way that statistical method for extracting the signal works is really different And I think it improves the accuracy and the temporal precision of fMRI Thank you Thanks I Because you need to have one slice one of each of the slice of the volume to be exactly the same time after The experimental design so first let's say you have 40 slice It means that you have to have a jittering of 40 the jittering And that means that the sampling of those slices doesn't I mean that that many something That's one that's one remark. The second remark is that it's in the same the sampling thing It's not like a you know only something in time as soon as the subject moves You have a 40 sampling problem Right, and that's that I haven't seen at all any of the I mean, that's a major problem That's what that's massive problem. So that may be completely so I mean the constraint on the experimental design and the Movement seem to be two major aspects that you're not discussing very much I think a general comment on that is that there's no free lunch I mean if you want to improve the temporal resolution you have to do something I mean is the same machine the same data. So something has to change Yes, you have to present more stimuli to do this. There's no way around that With respect to motion. Yeah, that's I mean, this is a problem in fMRI And and this Yeah, this is something we should look we are but I don't need to look at I mean it's it's I I mean I have looked at many issues such as motion correction and and Other types of preprocessing Ideally, I would like to do that with a simulated data set. So I know what's going on But it influences the analysis for sure. Yes Do you do the slice repositioning or how you call it before or after motion correction? I Did both I And it matters. That's a similar point that was made there And with these picture naming data if you if you look at the signal that is extracted Before motion correction after most in motion correction. There's a difference in the in the similarity of the signals. So Yes, there's it matters for sure and how do you know which one is better or you don't know You don't know. No, yeah, that's why I think we should do Simulated data sets where we control The ground through signal and we can see exactly see what is the influence of this. Yes Thank you very much. Thank you