 So the thing I'm going to do today is to basically give you a very quick and broad overview of brand imaging methods and some of the analysis that we're playing and some of the problems. So please interrupt me whenever you would like. It's going to be informal. It's not going to be very formal. And I probably have too many slides, so at some stage we will be skipping a bunch of that. So hopefully you learn something and again, interrupt me. So in the preparation of that course, I actually stole a lot of slides from many people on the web. So I just wanted to give credits now because otherwise it's going at the end of it. It's probably going to be lost. So all those people I've really taken some of their roles to present to you today. So I just wanted to thank them for that, for the work. So this is the outline. So I'm going to give you a very short overview of the acquisition technique. I think brand imaging you start by acquiring things and knowing exactly what you acquire is one of the key aspects. And I'm going to be a bit more in details in the MRI physics. How many of you are physicists here? Not none of them? Okay. None of you? Okay. So the MRI is just a beautiful machine, it's just a fantastic machine, I would just like to transmit a little bit of enthusiasm for that. And then I'll go a bit more on neuroscience aspect, can we average, can we do some population studying with brand imaging, what are the problems and so on? Then I'll go into anatomical information, diffusion information, then functional information, and hopefully the bulk of the course today will be on the functional aspect. And then maybe some more recent topic, like prediction, can you predict what the subject is doing? Can you predict a disease using brand imaging, this sort of thing, just laying out a couple of ideas on that side, but not going into more details. And the last part, which I may keep depending on times, is a little bit of a, is an important aspect, especially for teaching, is how do we make sure that our studies in your imaging, in brand imaging, are reproducible. And that's one problem in science and neuroscience, that we have to solve as a community. So let's start quickly. So you've seen probably yesterday, I'm sorry I couldn't be here, but you've seen probably the level of a microscopic level, at the cellular level, maybe you've seen some things at the corticon column level, and the thing I'm going to talk about is more like the macroscopic aspect of the brand, and its organization. So what are the major modalities, and just for a little bit of a historical reason for teaching purposes, I'm going to talk quickly about X-ray, I mean CT, which is one of the first modalities where you have a good view of the anatomy of the brand. I'm going to also give a very, very brief overview, everything will be a brief overview. So I'm going to stop that to say that it's a brief overview, but so I'm going to talk about the nuclear medicine images, and then MEGAG, and then MRI. So spatial and temperament ranges, this is a kind of classical slide that you may have seen already if you're interested in the program, but all those new imaging techniques have their spatial and temporal resolution somehow. So if you look at, for instance, fMRI, fMRI is the resolution around the second, let's say, I mean the actual physiological resolution may be a bit more than that, but depending exactly what you call resolution, temporal resolution, it may change. But basically it's the order of the second, it's a vascular sort of a modality. PETA is more like the order of several seconds a minute, EEG on the contrary has the resolution of a minute seconds, that's the rate at which you require data. And the spatial resolution, I mean you have this other axis, how well can you see the spatial aspect of the brain, and MRI can go very, very low in that. Like if you look at anatomical MRI, you can go from like 700 microns, and almost like sometimes if you have a very high field and the animal study, you can go very, very precisely on the anatomy of the brain. But obviously you have that, I mean having both the special and the temporal resolution is still a challenge in your imaging. And if you increase those things, you increase the amount of data and the potential problem that you have dealing with the analytics of the data. All right, so that's kind of a quick overview of like if you think about the modality in brain imaging, think about what are the actual new science that you're looking at, what are the temporal resolution, what's the spatial resolution. Those three things really should be in your mind when you're looking at a modality. So anatomical versus functional, and then this is like a very precise anatomy of the brain of a tumor, but what is really going on in a tumor's brain? Once you've got, let's say, the structure being the cables and so on, let's say you've got the CPU and the bus and all those things. What is the information being transmitted? What is the, so those are basically both modalities and both go hand in hand. And if you want to do a good study, you really have to think of both at the same time. But those modalities are more or less sort of dedicated to one aspect. If you look at X-ray, fluorescence, obviously, or CT, this is anatomical modality, ultrason, nuclear medicine is more like a functional modality. You'll be looking at the neurotransmission aspect. MEG is only a functional modality. You've got very, very little information on the structure. And one of the beauty of MRI is that it really has the potential to do both. And that's why it's such a major modality in brain imaging. So how does CT works? Well, you've got X-ray source, you look at the source, you turn around the objects and it's a tomographic modality. Basically, you're looking at slice by slice, tomos in Greek slice. And you reconstruct the information having the projection of all the objects around you. So when you turn, you get all the projection and you get the radon transform and you can reconstruct the inside of what you're seeing for the projection. And the major aspect I would like to emphasize is that it was actually possible and makes the thing real and usable is the actual FFT, the fast Fourier transform. There was a major sort of key element in these modalities to be useful. So the development in algorithm and prime mathematics really has a key. It's one of the things that will load those modalities to flourishing to be very useful. So that's a brain city, I mean nothing very specific about that. Nuclear medicine, it's almost the same principle except that instead of having the source of the radiation being outside of the object or outside of the brain or what you want to image, it's inside. You inject, for instance, in the case of a spectropec, you inject a radially ligand which is going to emit some radiation and what you're detecting is around this and you're detecting around the object radiation. And that's almost the same thing. Then you also have the tomographic slide by slice and you also have the aspect that you have to reconstruct what is inside the object having only the information from the outside. So again, you're going to use FFT, write and transform and apply mathematical rules. And that's what a PET scanner looks like. I think when Total Recall was filmed, the machine for Total Recall is very much like a PET machine. There was the new machine at that time and this is the sort of thing. And this is actually Bill Jaggers from Berkeley. So PET, as I said, is basically you have a bunch of detectors outside of the brain. You've got a radially ligand and you're measuring the, you know, basically the radially ligand is a positron emitting radially ligand and when the positron meets an electron in the matter, they sort of emit two photons at 512 kV. And that's the detection that you have. You're looking at all the detection of those photons, you know. And because you know that when there's a lot of activity there, then you are able, by looking at all the detection around the brain, you can reconstruct where it comes from. That's the principle. And what is useful mostly is oncology. It's very, very important for oncology because the tumor, you know, they have a higher activity and that's a very useful modality for that. It's potentially for, you know, diagnostic, you know, for the spatial distribution for drugs or like, you know, farmers are using PET quite a lot because they want to know what, where is the drug having an effect in the brain, for instance. Cognition, it's also a useful thing for, let's say you have an endogenous, a neurotransmitter, for instance, you know, you've got dopamine, you know. You've got the distribution of dopamine in the brain. And, you know, you may have a ligand that is actually displacing dopamine. It's just replacing dopamine. So you may be able to actually look at, you know, how much neurotransmission is going on by how much of displacement is occurring, you know, using PET. So it's a very, it's actually the technique to look at neurotransmission, which is not something which is easy to study, you know, in vivo, in the human subject. So that's, that's extremely useful for that. And, you know, constructors are actually doing now PET and CT machines. They have like two modalities in the same machine and it's very useful for co-registration and for just, you know, the economical quick aspect of having only one, one machine doing both things at the same time. So let's turn to EGMEG in a couple of slides. So this is because the PET requires more, so more, more hospitals have fMRI or have MR than PET, right? So it's, so I just have an idea about the, like, a number of scans done in different modalities or like a ratio, I mean. So what is more? It's difficult because MRI is used both for clinical and, I mean, I'm more like a study on the side of research and cognitive neuroscience aspect of that. But for, there is clearly MRI is a much more ubiquitous machine. It's much, I mean, like there are millions of scans being done every, you know, year. I hope I'm not saying the city number, but I think it's very, you know, it's a very, very high number anyway. And PET is more, it's more expensive. It requires some chemistry. It requires that you have, you are, you're able to label a ligand, you're able to, and then to, you know, and also it has some effects of the city, many of you, it's very controlled for the, for like a studying condition because it has some, a little bit of radiation like, you know, many of the tools that we're using have a little bit of radiation. It's a matter of whether there are no, no, no see radiation enough. I mean, you know. So it's a, but so it's, it's, it's much less, you know, spread machine because of those, those things. But in oncology, it's very, very much used. Spectant and PET are both very much used in oncology. Yeah? The PET scanners tend to be located in hospitals where they do a lot of oncology. So for patients who, and it's usually often used after the initial diagnosis, people don't use it to make initial diagnosis, it's usually used to look for metastases and other things. So the fact that it's more restricted probably isn't that serious in terms of it's widespread medical. And the main lag and for that is FDG, like you're looking at how much consumption of glucose using the fluor. So EG MEG, that's very widespread, I mean EG is extremely widespread. It's, I mean the clinical aspect of that is epilepsy mostly, but also other things. And that's, but the research kind of tool is more, I mean both are used for research, but the MEG is more like the interesting machine. It's beautiful, you've got a very elegant hat on the head. And it's basically measuring the little current that your brain is the electric current that the neurons are doing. So let me just have a quick recall of that. So the spikes are not really the thing that we can measure with EG MEG. So EG is the measure of the electric current while the MEG is the measure of the magnetic aspect of the current. So you know that each time you have a little current you have a little magnetic field around the current, right? So EG is measuring the actual electrical aspect and MEG is measuring the magnetic aspect. And it's not really the presynaptic spike that is measured, it's actually the post-synaptics which is, you know, has a longer time frame, so about tens of milliseconds. And that's where, you know, there's enough of that can be accumulated and average across, you know, enough dendrites such that you can measure something because it's obviously such a very little signal that you're trying to measure. So this is so EG and MEG kind of recording. So the post-synaptic currents that we're looking at, it's EG is a very, you know, old technique, 1930, MEG, and I'm not exactly sure. Last slide. Do you want the previous slide? Yeah? I mean, sir, is it current or potential, that figure? Oh, it's both, I mean, it's the current and the potential. It's a- So that figure is about potential. Yeah. But you know, you're measuring the potential of the current and you're measuring the magnetic aspect of that current. So in MEG, with MEG, you're measuring the magnetic aspect of that current. And so the temporal resolution is excellent. I mean, and you're measuring that, you know, like every millisecond or so. And it's a much more difficult thing to measure the magnetic aspect. You've got squids, which are, you know, like a superconducting helium and it's a much more difficult physical problem. But also it gives you a little bit more of a spatial resolution. The spreading of the potential is much more that with the skull and the skin and all those things. The currents are kind of, you know, what you're measuring is very much spread. While in MEG, you've got a better sort of a localization aspect. So, and you know, and I'll just put that slide because it's a, you know, it's a classic representation. So you look at the, and all those slides are from Alex 1-4 from, you know, Paris that was very kind coming, gave me those slides. So the presentation is basically that you've got, at the end of the day, you've got a big matrix, which has in one direction the temporal aspect and the other direction the spatial aspect. So here, for instance, you've got one column here is the number of captors that you have, you know, either in MEG or MEG. And one line would be like for one of these captors, what's the temporal sort of a, you know, and you know, a lot of the analysis method are actually going to, you know, look at that matrix and try to tease apart what are the sources, what are the, how do you summarize this information, how do you, and so on. So it's a useful kind of mental representation that you've got, you know, temporal, spatial and that's going to be the same for fMRI, for instance. And the, I mean, it could be temporal, it could be also subjects, like you know, it could stack subjects, information if you want to give a group study or something like that. I've got to watch the time because I don't know. Okay. Right. So what can you do with a MEG, so cognitive studies, like looking especially at temporal aspect of, you know, your cognition, diagnosis for epilepsy, that's one of the big thing. Grand computer interface, you may have heard those applications that you're trying to from, you know, the current, I mean usually EG of course, but how can you from that command a computer or an apparatus just without, with your, with the mental activity that you're generating. That can be very useful for handicaps and things like that. And the name of the game in EG MEG is actually you measuring on the surface, like with electrodes or squids, the electrical or the magnetic activity. How do you know where this electric or magnetic activity comes from? Where is the actual electrical source that is generating that signal that you're measuring? And that's a very complex problem because it's an impulse problem. You've got many more possibilities of sources than you can accept many, an infinite number of possibilities of sources, of organization and given one measurement on the head. So you have to impose some constraints and there are clever algorithms to sort of try to find where are the sources and their temporal activity. So that's one big sort of an analytic problem that the MEG EG communities is facing. So let me turn to MRI, which is my favorite, the modality I'm working mostly with. And I just, I mean, the first time is basically that MRI contrary to the other modalities is not basically one, it's one machine, but it's many modalities. And that's the strength of it really, is that you're able with the same machine, depending on how you program it, to look at very variety of neuroscience aspects of the brain. And that's the key aspect of the MRI and that's why it's so useful and such beautiful machine. So here I'm going to talk a bit about diffusion, anatomical information and functional information. Just to end up with a EG and MRI, you can actually try to record both EG inside the magnet, inside the MRI. And I've just wanted to put that slide because notice it's a big deal as well to try to have both the spatial information and the temporal information. And if MRI is giving you better spatial information, EG is giving you better temporal information, can you get both at the same time by measuring conjuntly in the same subject at the same time, simultaneously? So those are interesting studies and pose a lot of good physical challenges because in terms of recording very, very tiny potential in a huge magnet, you get a lot of interesting problems. One thing I want to mention which is not really an imaging modality but it's a very important and interesting modality and it's linked somehow to MEG. It's kind of the opposite of MEG. MEG you're just recording the magnetic field. TMS, which is a trans-magnetic stimulation, is actually you have a little magnetic magnet and you inject some current, changing current and therefore you're generating a changing magnetic field and this magnetic field is actually going to induce some current in your neurons. So this coil is going to, you can stimulate a part of the brain using TMS and that is mostly used for sort of a shutting down a region temporarily somehow. You're disturbing the activity of a specific region. Oh, you showed a video of that. Excellent. Okay, so I'm sorry, I missed that, but yeah. It's a, I mean, in terms of ethics, it really should be considered as a medicine somehow. You are actually impacting the brain and that's a, it's not a recording thing. You are actually inducing something so it has to be very carefully monitored. All right, so MRI physics, I'm going to give you a little bit of a primer on that. Again, it's going to be very rough and hopefully you'll get basic ideas and some maybe enough to sort of go back to the web and look at some things that if you're interested. So what's a scanner? What's an MRI scanner? The MRI scanner is first of all a huge big magnet like a free Tesla machine is like the magnetic earth is about 10 to minus five Tesla. So it's 100,000 more than the magnetic earth. So it's a huge magnet. Anything ferromagnetic that you put there is going to fly, it's going to be very dangerous because it's, and that's a, and then it's got more magnets somehow so that you, this big magnet, you can change the field depending on the location that we call that the gradient. And it's got cores that can induce radio frequencies. So if you change quickly, you can induce with magnets to radio frequencies. So all those things are within the same sort of big machine. And it's got, and there's this code here. It's just like receiving the magnetic, the changing magnetic field around the head. So that's what is in one word in three words, like an MRI machine. And to get a big magnet, to get a very high field, you need cryogeny, you need to have a superconducting currents. And so you get, you need to have all this. So magnetic is always on the flow. It has a non-resisting current going to the magnet and it's always on. So you have to be careful. You never shut down the MRI. It's a, unless there's a problem, it's called a quench and you have to do a better. So what's the principle of that, of the acquisition? It's not an easy, I mean, I could tell you the principle of CT, basically, you know, couple of words, you have a, you know, radiation and you look at you, how much the radiation is attenuated for the object and you know that, and that's it. And you, you know, it's not like a complex, you know, thing to understand. MRI is a bit, you know, requires a bit more like a, you know, physics and is more involved in terms of, and that's what, you know, it's kind of fascinating. So MRI will look at, you know, each of you, let's say it takes you the water molecule that you have, you know, maybe there's a lot of water in, you know, in all our body and brain as well. And the atom of the, the hydrogen atom has a proton that is actually acting a little bit of a spin as a magnetic spin. So basically the, those things are, have some magnetic aspect and they turn around themselves. And so, and all those, you know, those atoms and, you know, the, the nuclear of those atoms, they act a little bit like small magnetic spin, right? And that's the, that's the principle of it. So you have to go very deep into physics to sort of know the origin of the signal. And so if you put those things, if you put a head or like an object in the, in the magnet, those little magnets are going to align with the big, you know, B zero field, right? So they're going to go all in the same direction. And they're not only going to go in the same direction, they're like, as I said, they lack spin. And therefore they, if you take a spin in the, sort of the, on Earth, if you spin it, you'll see that the axis of the spin is going to have a precession movement, right? And those, you know, protons are actually doing the same thing. They act, you know, within this big field, they act like small things and they have a precession movement. So that's the, that's the origin of the, of the signal. So it's a bit very deep, like in the, you know, the physics, the matter, right, the physical matter. And so it's interesting. So if you now put a radio frequency at the same frequency of the, of the, of this precession movement, then you can, you're going to align all those spin and you're going to align all those little magnets somehow. And that, if you do that at the, at the frequency of the precession, you're going to give more and more energy and you have the resonance effect. So all those things are going to align and they're going to synchronize with a radio frequency and you're going to do that and, you know, and then, you know, you got a lot of energy and then you stop and then you listen. How does that go back? And so you're, what you're looking at is just, you know, the, you know, after resonance effect, you know, how the relaxation aspect of the, of the, and these relaxation aspects will have different temporal and values and, you know, and a characteristic depending on whether you're in a gray matter or whether you're in a CSF or you, you know, so it depends on the, what you're looking at. So I don't know if you know, you could get that, but I think that's the principle and it's a beautiful sort of a, like a very deep into the matter sort of a thing. So you've got a very, you know, technological apparatus that look at very deep physical thing and it's, it's interesting. So to summarize this, what I've said, you know, if you look at just the average, you know, magnetic field that is, you know, sort of the thing that you're interested, you're going to, you know, do the impression, so do the radio frequency, get the resonance effect and then you're going to look at how this resonance effect is going to sort of grow back, you know, towards the B zero value and decrease towards the transverse plane. So, so if I do this, this is just one, my one, some summary of the, of the magnetic, you know, moment that I'm looking at. And so I'm just putting some energy with the radio frequency, okay? I've got the resonance effect and then I'm listening and it's going back. So it's going back two ways. It's growing up again along the B zero and it's growing down along the, you know, the transverse axis, right? So that's the, what this little thing is telling you. And if you just take the rotating some sort of, you know, coordinate system, you can look at it as, you know, going down to the transverse plane and then going up again. And what the, the T1 and T2 sort of things are, you know, when you're saying, you know, you're looking at MRI and the guy will tell you, this is a T1 MRI or this is a T2 MRI, weighted MRI. And those, those are the, the time constant by which, you know, they grow up again. So T1 is the time constant by which how much, I mean, how the thing is growing up towards the B zero value and T2 is the time constant, how it goes down on the transverse aspect. So those are the two time constants. I'm going to skip that. So this is just, you know, the curves that you look at the T1, how the T1 is growing up again and this growing up again is different depending on where you are in the brain and the going down, you know, along the transverse plane is, you know, it has different time constant depending on where you are as well. So this is the T1 versus T2, weighted MRI. So this is how you distinguish matter in, you know, by looking at the time constant of, you know, relaxation aspect of the magnetic information. And the interesting things that the biology is going to kick in there, if you have more oxygenated blood flow somewhere, you know, it turns out that the oxymoglobin is actually diamagnetic and the deoxymoglobin is paramagnetic. And so basically the local magnetic field of the oxy and the deoxymoglobin are different and therefore you're going to be able to measure something which is related to how much oxy or deoxymoglobin you have in the brain. And that's where the biology starts to sort of mix up with the physics aspect and that's why MRI is so, you know, like a versatile because, you know, you can look at those things. Yeah, so the oxymoglobin have like a less disturbing effect on the magnetic field than the deoxymoglobin. So everything I told you, I told you we're looking at the amount of like magnetization that is coming back and so on but where does this magnetization come from? I mean, I've told you there's an antenna that records, you know, measure that thing that is coming in this relaxation effect but, you know, how do I localize thing? How do I know that, you know, this is the one in that space, in a particular place in the brain versus, you know, this one? I mean, I haven't told you anything about that. For the moment, you just have one object that gives one sort of a signal, right? So that's, again, magic. The precession movement that I was talking the actual, you know, frequency of the precession is depending on the B zero. So if you have a high field, it's going to process very quickly. If you have a lower field, it's going to process you know, slower, right? And that's the Larmor frequency. That's the discovery of Joseph Larmor in the late 19th century. And so what we're going to do in the magnet is put some gradients such that we're going to have a slightly more a higher field in one hand, one end of the brain and then decreasing, you know, like a lower field in the other end of the brain. And then we're going to listen to this whole thing and we're going to know that the higher frequencies are going to be on the left of the brain and the lower frequency are going to be on the right of the brain. So the principle of the localization is really like, you know, you're changing the frequency and you're looking at, you know, just measuring the whole thing and then you're teasing apart what is higher frequency and therefore where is in the, it's like a piano. If you have a high note and you know it's on the right of the piano, if you have a low note, you know it's on the left of the piano. And if you have several at the same time, you can do again for your analysis and looking at, you know, how much, you know, the low note had and how much the high note had. So I'm not going to the details because that's give you the localization aspect on one axis. We have another axis, which is the X, let's say this is X and you have the another axis, which is Y. So you have to do gradients on the other axis as well. So you have like a, and so on. And there's a lot of tricks and it's a fascinating area. And then the other thing is the, how do you know on the Z direction, when in the Z direction, you can actually do the same trick but at the excitation level. So if you have a gradient in the Z direction, you know you're going to excite only one specific frequency depending on the Z direction. If you know change the Z direction frequencies, the magnetic field. So just to summarize this, I mean, I'm taking too much time for that because I find it fascinating, but just to summarize the way you localize things is actually by looking at the, varying the gradients, the magnetic field with the, with space. And that's just an illustration of what I said. You know, we've got high frequencies on the right of the brain, low frequencies on the left of the brain and you're measuring that and you're taking Fourier transform and that's it. And that's actually the actual recording that you get from a slice in an MRI and you're taking the inverse Fourier transform. What you're looking at is the actual, the frequency magnitude for, you know, so these are the low frequencies or actually the kind of the shape of the brain, things like that. And you know, on the outside, you've got the high frequencies which are going to give you the details. And you're taking a Fourier transform. It just, so yes, again, you've got like the, those tools that are going to be super useful for the, so I'm going to keep that, but basically, you know, using those tools and all those things, you can look at the anatomy of the brain, gray, white matter and, you know, skull and all those things. And that's the T1 weighted. You can look at something. I mean, we're going to, from functional MRI, we're going to use that T2 weighted images. And I'm not going to Y, but basically that's another T2 weighted could be very well in anatomy code images as well, right? It's just, it doesn't, but for functional MRI, it's going to be the one we're going to use. Just want to show you, this is the B0 map. So what is the magnetic field when you've put a head in the scanner? And because of the interface with the air and the tissue and all those things, you're disturbing the magnetic field of the scanner, which has to be very, very homogeneous to be a good MRI machine for recording, you know, the images. And one of the problems that there is, that's especially with MRI, you're, because of this disturbance, you've got distortions or lost of signals. There's a lot of artifacts. One thing that you have always to recall is that, you know, anything that is ferromagnetic in the brain is a problem and can be very dangerous for the, you know, if you have a, so this specific brain, you shouldn't put in the scanner. I don't know how that happened, but... It's a... Yeah, there are accidents, yeah. So many contrasts. So I was telling you, you can look at, you know, anatomy function, blood velocity, blood flow, diffusion, which I'm going to go into in a moment, and more in the future. I mean, you know, like biologists working with physicists in this area are actually discovering and, you know, developing new contrasts to look at new things and so on. So it's a beautiful mix of biology and physics that is... And again, the machine itself, if you think of it, is actually more amazing than the space shuttle, I think. I mean, it's... Okay, so this is just to show you the, like, you know, this is, you know, an anatomical MRI and this is the fMRI that we're going to look at and you see, you know, the loss of signals, like, close to the air tissue interfaces. So let's switch gear and let's say that we have a machine to look at the anatomy of the brain. And the question is, how do we do studies across population? How do I take you to this classroom and then, you know, measure something on the functional aspect or the anatomical aspect of your brain and then have a, like, you know, an information on the population that is the sampling of the population that you are. And the community has soon, I mean, soonish, converged towards the idea that we should have an atlas of the brain and one of the first atlas was actually developed by a functional surgeon who actually needed to go and, you know, localize the deep structure, the telamy, the codec, those deep structures in the anatomy of the brain. So he thought, okay, I'm going to sort of, you know, have an XYZ coordinate system. I'm going to put the anterior commissure and the posterior commissure at those places. I'm going to have, like, all the brains aligned with the same orientation with respect to those little structures. I'm going to have the box of those brains, like, you know, in the same box. He had a little bit more complex, you know, way of doing that. But, you know, basically that's the, and then if I am at the X equal two and Z equal three and Y equal minus five, I know in which structure I'm going to be roughly. And that's the roughly which is. It works quite well for the internal structure. But as soon as you go into the cortex, that's a little bit of a never-matter and we're going to describe that. So that's the name of the game of what we call special normalization or, like, you know, the imaging, in fact, the image processing community is calling that core registration or registration sometime. And so basically you're taking one anatomy of a brain and then you're deforming to look like a template brain. And that's, you know, that's the principle. So there was a lot of development of standard templates, like the Montreal Neurotracheal Institute has the most famous templates of the brain that is, you know. And templates can be in several modalities. You can have template of brain using, you know, T1, T2 star, you know, pet machine have a template of a brain and so on. The idea is that you take all your subjects and you deform their specific image to look like the template. And it's a good idea possibly to use several of those modalities to do this transform. And there's a lot of problems. There's the technical problems, which is basically how much do you, do I morph? I mean, what, how much do you feel more thick, you know, variation? I mean, how much, how much degrees of freedom do I put in this transformation? I mean, if I do linear transformation, okay, that's fine. I've got like maybe six degrees of freedom, maybe plus plus three or plus, you know, six, depending on if you want to share or, you know. But if I do, you know, a more thing, like, you know, like, you know, I can morph anything to anything if I have enough degrees of freedom. So how much of that do I do? And sometimes the, you know, like you see some deformation that are not biologically plausible because of this problem. And sometimes the patients or the brain that I'm trying to morph is actually missing piece compared to the template, and that's another problem and so on. It's, despite all those kind of, I'm going to go to the neuroscience problem, but despite of those problems, it's very useful to sort of have a rough idea of where things are. So if you have a labeling of the structure of the, in the template space, and if you have done this more thing, then you have a labeling of the structure and your actual, the brain you're looking at. And that's basically what the thing is doing. But there's a lot of quirks and difficulties with that. So one of the main question is really is what information do you use to go from your brain to the template? What is the actual thing you're trying to match? And basically what you're trying to match is a big contrast. Like if you get the border of the brain, that's a big contrast. And you're going to sort of make sure that the border of the brain has the same place. So, and again, if things are not too valuable, if you have a very standard number of a structure, that's fine. You can actually sort of try to morph thing and that sounds okay. If you have variable structures and you don't have a one-to-one mapping, that's not fine at all. We don't know what we're doing. And that's what the, one of the problem is. So let me just mention this because I think for neuro-informatics sort of a community that's a new thing to study. So how do you validate that you're doing the right thing by this morphing? What's the best algorithm to do this morphing towards a template? And there's a beautiful study by Arno Klein that has a look at that. He had a set of manually labeled brain and then he looked at how much those labels after morphing were matching the template labels or those kind of ideas. And it's, you've got many, many algorithms to sort of, and development to sort of people trying to do the best morphing aspect. And you've got very, very few validation studies of what are we doing the right thing with those morphing. And that's one of the things that I would like to emphasize. Those validation studies, they are the key aspect for that problem. And there's several labeling in the community. Like people have labeled the structure and the brain in different ways. And there's no, and that's also one of the things that the neuro-informatics community should work on is to how do we get consensus on those things. And there's a beautiful study by Bolland that showed that, you know, we're looking at eight different atlases. You've got very, you know, got your serious, you know, matching problems. That, you know, if you look at the temporal gyrus, you know, in one atlas, it's not going to be the same. Even if it's the same template, it's not going to be the same that in another atlas. And the sort of the overlap of those structures across atlases is a real problem. So we're doing very, very rough thing, very, very rough because, you know, we don't actually know whether we're doing the right more thing. And then we don't actually know what is the right labeling of the template. And so that's to tell you how much rough things are. And this is the reality of the variability of our cortex. And if you look at, so those are three cortices with like a labeling of some sulca and it's a beautiful work by Kashi and all the, my colleague at NeuroSpin. It's just that if you look at the sulco gyrus variability across brain, you have a very big problem to know how you're going to match those two things. That's how do you match two subjects that don't have the same sulchi and the same gyri? I mean, let's say you have a, there's a very famous, like at the, in the cingulate area, for instance, some of us have two sulchi and some of us have one sulchi. And you can't match that. There's no one-to-one mapping between those two structure, right? And so that's the problem that the Neuroscience aspect, you know, apart from the sort of the image processing aspect, that's the Neuroscience aspect of the problem. And then how do we, you know, bring in other information, which is not the sulco gyrus information, but the, maybe decide to architectonic information. And there's a lot of excellent work by the Zilis lab, looking at, you know, post-mortem brain and seeing what's the situatectomy of the brain and how do we map that. And, you know, there's a lot of good work and interesting work, and so can we have a better mapping of those situatectonic map if we have a registration which is going to be on the surface of the cortex using the information of the big sulco gyrus and these sort of things. That's a good study by Fischer a few years ago that if you're interested in that problem. And basically Fischer showed that, you know, looking at doing this special normalization using the deep sulco gyrus is actually giving you a smaller variability of the situatectonic information for instance. Right, but still, you know, this is what the tool that we are using for us and that's how much things improve with the processing that's a John Ashburner sort of work on the templates and this is like the average of 452, you know, anatomy of subject anatomy. And you see that there's very little, you know, sort of a gyrus and sulco information in this average because, I mean, there's some information. There's, you know, like there's a smooth information and that's more like using a better algorithm to match things, you know, given the fact that with the validation of our own client kind of, you know, constrained somehow and you see that the sulcus and the anatomical information is getting better and better with those things. So conclusion for the, how do we average brain? Basically, we don't know how to average brain. I mean, I guess the only thing would be like you have a neuroanatomist that is looking at, you know, what specific area can I average this area with this area of this subject and this subject that's just practically impossible to do, you know, in most of the cases. So what are the reliable information that we find across subjects that we want to match? That's one of the big neuroscience question and neuroscience neuroinformatics question. How do we validate? So how do we sort of include probabilistic information? I mean, I've talked about structure and things that we know there, but you know, what's the, what about the probabilistic information that we can get? How do we update templates or to update at classes? Those things are like, you know, there's no versioning, good versioning system and all you like, so there's all those things and, you know, how do we deal with both landmarks and maps? We know that there are, you know, sort of a maps of like a continuous information. How do we, I'm not talking probabilistic, but also just continuous information. How do we map those things? Despite this, there's a, I mean, I'll tell you later on if I have time, the Jungler story because that's a beautiful study, but there are many bad examples. I'll tell you a bad example because it's always more fun. One of the first study in brain imaging was a study of a, with PET, and they were looking at the fear and anxiety. They were studying anxiety. They wanted to know what are the brain structure involved in like, in processing, you know, anxiety. And they had those subjects in the scanner and so on, and they had something, some stimulus that were generating anxiety, right? And they wanted to compare with a stimulus that was not generating anxiety, and that was a study. And it was a nature paper, or science, maybe I can't remember, but it was nature or science. So one of the best papers, they looked at, you know, beautiful brain, map of the brain, like a red, you know, color, you know, area, and so on. And then, you know, they looked at it a bit more carefully. I mean, you always want to publish quickly in science and nature. They were looking more carefully, and what happened is that the subjects were clenching their teeth when they were anxious, and then the muscles had a lot of blood flow going in, and because the spatial normalization to the template was bad, they thought that the activity that was in the temporal lobe was actually in the muscle of the subject, right? So that's a, you know, it's a long, it's a study that, you know, think it's 92, three, I mean, one of the first studies, or it's a, you know, we've passed that, the community, but you know, errors happen a lot, yeah. And that's a special, a famous special organization error. Diffusion imaging, I'm going way too slowly, I'm sorry about that. That's a, okay, well, yeah. That's how it goes. Diffusion imaging, that's also a fantastic thing from MRI, because that's the only in vivo modality that can map how your brain regions are actually connected. I mean, there's no, and even if you postmortem, if you open a brain and if you look at me, how do I map? How do I know which area is connected to which area? It's just, you know, even with the best, you know, neurosurgeon, you know, trying to do that, it's extremely difficult because it's a 3D object that you can't actually, you know, investigate easily. So, diffusion imaging is the MRI modality that allows you to look at how much the water is diffusing in one direction. And when you've got a fiber track, you know, connecting, you know, two brain regions, you've got a little bit of more diffusion in the direction of the fiber track than across the fiber track, right? And that's the principle of it, and I'm going not to go into that. Also, should mention that diffusion imaging is used massively for stroke. So it's a very good modality to look at, you know, if you have the stroke, you know, how big is the impact and, you know, it has some effect on, you know, how you can treat the patient. I'm not showing you the way this is measured with MRI. I'm sure you can find that on the web. But basically, you measure the diffusion, you know, in several directions. So in this instance, we measure the diffusion of water in six directions. So this is in the X direction. And you see here, for instance, the corpus calluses is which is linked, a big set of fiber tracks linking the two hemispheres of the brain is actually having a high value because there's a lot of diffusion, you know, going into the X direction between the two hemispheres. And that's one, you know, just to show you how it works. And then, you know, if you measure that in many directions, you can try to reconstruct, you know, how, you know, the brain regions are connected. And that's how this is illustrating. There's a lot of problems, which you will always be skeptical as scientists. You know, if you look at something, that's the problem of crossing. So the fiber tracks are going this way. You know, the information that you get, if you just measure in one, you know, in two few directions, you get something which is not showing you the crossing. So you don't actually, there's the kissing problem. I'm not saying that kissing is a problem in general, which I'm thinking that, you know, it's when two fiber tracks are going like that and then like that, so it's how do you have the resolution to go into that and so on. So solution is to both increase the resolution, spatial resolution, both in space, having, you know, smaller voxels, but also in the angular direction, you know, measuring more directions and better tracking algorithms. Yeah, so that's, for instance, the kind of a diffusion sort of a probability, what's the probability of the diffusion in space. So each voxel, that's an interesting sort of modality. Each voxel can be a little 3D image itself, where for each voxel, you have the direction of how much diffusion you have in each. And then you have the problem that you have also with T1. How do you average those things? How do you construct something which is for an average population? And that's the same sort of problem you get. So the big tracks, you can probably sort of map an average, easily the kind of the smaller tracks where you got higher individual variability. We don't know what to do with that. And then there's a lot of people who are interested in grad theory. And then you, so let's say you have, you know, an idea of how the rich region is connected to each region. You can look at, you know, a parcelation of the brain and then, you know, see how much of this, you know, a track is going from one parcel to another parcel. What's the connection? And then have a kind of a connectivity matrix that you can threshold and then look at the graph that this is producing. And then many people are actually interested in whether this graph has properties that are interesting for the brain activity, whether they are, like, is it a modular thing? Is it a small world graph? What's the efficiency of the graph and all those things? And those things can be different between populations of patient versus normal, or, you know, depending on, you know, and also using, you know, natural neurology, you know. All right, so let me talk a bit of a functional MRI. So functional MRI is, so I told you, you know, that you're basically going to look at the activity of the brain but at the physiological level of the blood response because, you know, this is what we can measure. And so what is the research in functional MRI? So first of all, people are interested in mapping what brain regions are doing, what connective function. Also, the functional connection, even if I have a track, is it used during a specific task? You know, let's say I have a track between two brain regions, but how much information is going from one brain region to the other? And there's a lot of a neurological or psychiatric sort of possibilities of, you know, whether, you know, for studying, understanding the mechanism of some of the disease, possibly following up some treatment and so on. The only real clinical application that is at the moment used for MRI is the probability replacement for the water test that is, you know, telling you whether your language is mostly on the left hemisphere or versus, you know, on both hemisphere or on the right hemisphere. There's a, as you see, there's a lot of functional variability between subjects. I mean, like just knowing that your language can be mostly on, I mean, my language is mostly on my left hemisphere, but it may not be the case for all of us here. It's a big functional variation, right? It's a... Let me just backtrack a little bit because that's an interesting sort of a, when you know, you see probably in the magazines and newspaper, like, oh, this is where the processing of a mathematics happens in the brain and we found out, you know, where, you know, what is the region, where is the region when you think of your grandmother, you know, and so on. And, you know, remember, maybe some of you, if you've studied a bit of a history of science, you remember the phrenology, you know, sort of a, and the phrenology where people are looking at, you know, your brain sort of a shape and when they had like a, you know, a whole or a little bit of like a variation in the shape of the brain, they would tell you, oh, you have a very strong, more, you know, instinct or you have, you know, you're very lazy here or, you know, you must be very good at mathematics or, you know, like, you know, by looking at the, you know, just the shape of the brain, of the skull, right? And not of the brain, the skull. And the critic of fMRI is often that we are doing by the same thing, you know, we just have a better tool but we're not understanding how things are functioning by just telling you where they are and there's a very famous analogy by Jerry Fodor who's basically saying, if you look at the motor of a car, you know, knowing where the carburetor is doesn't tell you anything of what the carburetor is doing in the motor of a car. And that's the, you know, the criticism and the very fair criticism that the, you know, this community is facing. And there are obviously good response to that criticism which I'm not going to, you know, talk about better. I have a question to the speaker. I remember that from last year. So why is this guy exactly put into the oven? I mean, what's the relation to phrenology? I guess this is like a MRI kind of machine of the time. And this is what they, you know, this is the machine that would, you know, look at all your thoughts and, you know, and. Ah, okay. Yeah. It doesn't really alleviate your fears of going to the doctor. Yeah, no. So, yeah, I was going to tell you a bit more in the history of how, you know, the bold effect was discovered. And it's an interesting story because it's basically a failed experiment that showed the bold effect. It was basically a rat in a scanner who died, which died, sorry. It's a, and they realized that the images after the rat died were different, you know, of a contrast than before. And they realized that it was the blood flow that was the, and that's how fMRI actually was born again. And it was in the first kind of a study was like more in 1992. So it's a 20 years old technique. One of the thing which is very sort of a difficult and interesting thing that has been done is to link the electrophysiology and the bold signal. So again, the bold signal is basically mostly the oxygenation signal, but it's got some, it's got some blood volume, blood oxygenation, you know, sort of a mixture of those things. So it's a bit of a complex physiological sort of a signal to interpret. But so to understand a bit more what was happening, Logothetis and the team actually went to put a monkey, open the brain, put an electrode, put that sort of thing in the MRI machine and then measure both the bold signal and the electrophysiology at the same time, which is a very difficult experiment. And you see that the local feed potential sort of electrophysiological measure are kind of like at this time, and this is during the stimulus, and this is the kind of signal that you get with bold, which is a much more sluggish delayed information. So it's not at the neuronal level, it's more like the, you know, when a set of neurons actually needs energy, they require oxygen and glucose basically, and that's where the blood flow is coming in and that's this overflow of oxygenated blood flow that we are seeing in with bold, yeah. So I've been asking about this bold signal two times already in the previous courses and I've been wondering the source of it and now I've done my homework and I know where it comes from. Excellent, you can explain. So I will explain it now and I think if you really dig into the references, you will find it. So it will actually show the activity of astrocytes together with vessels and this explains why it has such a slow timescale and this would be really important to emphasize for students here because I see many projects going on in the world where they model based on this bold signal and they try to fit the neuronal network activity to this signal. It's like bullshit to me. It's a complicated thing to do. I'm skipping over that because it's a whole field in itself to understand where the bold signal and what is the bold signal exactly. It's got several things. So the neuronal sort of origin is actually a truly a mixture of like neurons asking the astrocytes to supply. So basically neurons talking to the astrocytes, the astrocytes talking to the arterioles and then all the hemodynamic aspect. The only thing I think you should sort of be very clear it is not a neuronal sort of electrical physiological. It's a very sluggish, blood flow aspect. So that's the thing I wanted to. And we have very bad models for that. So this is the actual recording in this region for visual stimulation and this is the model that is usually used to model that thing. So it's a complex thing to model and there's a lot of work on that. And basically people are modeling it if you have several things coming, you know, stimuli for the subjects, then the question is whether those, is this signal actually linear? Is it, can you just use a linear model for, is it additive? And basically it's probably not in all ranges. It is certainly not in all situations but in many situations it's a rough, goodish approximation and that's what the study has shown that if you take three stimulus versus two stimulus versus one stimulus and you can see that there's kind of an additive effect of the ball signal in this specific circumstances. So how do you process that? So the fMRI acquisition is just like a brand volume every two or three seconds. The brand, the volume is quite quickly for an MRI machine so it has a very, kind of a two or three of, let's say three millimeter cube resolution. So it's a three, you know, it's a big, it's a number, it's a great number of neurons in one voxel, you know. And again, so you have to sort of do or the subject is moving during the scanner so you have to correct for the movement that you know these things that you're crying and then you know you're crying slice by slice so you're not crying everything at one time so you have to sort of correct for that slice thing and then you have to sort of co-register your functional signal to your anatomical signal and then you have to find out or depending which order you do things you have to find out which regions are actually involved in the stimulus that you've presented to the subject during the acquisition. So that's, and as I said, there's a lot of artifacts subject is moving and we can't correct properly for that. We can't even estimate properly the movement because it's a slice by slice, you know, acquisition and the subject is not moving like after one brain has been acquired, it's moving all over the place. So that's an example of movement artifacts that are, and then there's all the modeling aspect which is how do you, so we know exactly when, let's say we have a stimulus that is occurring at that time in the scanner and we've got a second stimulus occurring at that time and we want to know what's the difference of the bold response between those two stimulus and so on to the stimuli. But then you have to construct a model of what is happening and then do the difference and so on and this is generous, I mean, in general, this is done with general linear models. So it's just linear regression, we're just assuming generality and go along with that. And then because you're doing that, it's voxel by voxel, you're doing that for, let's say, 30,000 voxel, 50,000 voxel, depending on the resolution and you've got a massive statistical problem of how do you correct for the fact that you've done so many tests for so many. So there's a classic, and my background is a bit more in statistics, there's a multiple comparison problem that you have to deal with. So how do you correct for the fact that you're testing 50,000 voxels? You have to increase the risk, the type one error level. And this is one of the classic experiments where you look at the visual field and the eccentricity and the rotating stimuli and you can reconstruct the functional areas, V1, V2, and using those stimuli and a clever modeling of that. That's a, it's a beautiful error. And then as I said, you've got several subjects doing roughly the same thing in the scanner and despite what I've told you about the anatomical differences, roughly there will be like some regions that are going to be relatively similar across subjects. And then you have to sort of average that across, or combine that across subjects and there's a lot of statistical models to do that as well, mixed effect model and so on. And the, and more interesting I think is, can we have like a model that can actually be instantiated slightly differently across subjects? So you can sort of, so how do you build, how do you build an average with things that are not exactly at the same place? And, you know, and that's a, there's a lot of a good work done by my colleague, Bertrand Thurion, on these sort of issues. It's again, it's used often by clinic, in clinical applications, I see this is an example where you have a population of patients and then a population of control and you've got two conditions for each population. You've got let's say, you know, control condition and a placebo condition and a drug condition. And then, you know, the, you know, the farmers are very interested to know whether the drug may have an effect and a differential effect on the normal, on the control population versus on the patient population. And, you know, and they, they sometimes do ephemeralized studies to see that. This is an example. So, what am I? Okay, so I've got 15 minutes left. Okay, so I'm going to tell you very, very quickly on the end of fMRI connectivity and now that you have a little bit better idea of what we're talking about, it's going to be, you know, reasonable. So the question, I mean, most of the early work in fMRI was done on localization of a thing and with the caveat of the neophrenology that I was talking about. There's also the aspect of functional integration. How do regions talk to each other? You know, with a quote. And it's not, it is not like a, you know, a recent question, like in the, you know, in the 18th, 19th century that you had scientists working on animal models, you know, basically lesioning some part of the brain of the animals and some of them were thinking, oh, when I lesion that, you know, the animal can't do this, can't function in that area anymore. And some of them were saying, oh, you know, I can't remove a big part of the brain and I can't see any behavioral effect. So, and the model behind that was that, whether we've got specialized region doing something very specific in the brain or not, and you have more like an integrative, you need a number of regions to work on something too. And the answer, obviously, is that you have both some specialization and some, you know, integration. But there was a lot of, there's a lot of, you know, work now, and most of the work that was done to try to find the localization aspect has now moved to another area, which is more how do we, you know, estimate the networks in the brain that, you know, what the, and that's a, it has moved a lot with the, an idea which was very simple. It was a bad, bizarre idea that, you know, if you look at, you know, if you put a subject without any stimuli, stimulus in the scanner, and you look at, you know, how your functional signal in one part of the brain is going to be correlated to another functional signal in another part of the brain. And if this correlation is very high, then you think that, you know, there's no stimulus specifically, so it's just like the ongoing thought process of a subject, and then, you know, and by default, those areas are kind of connected and therefore they share some information. That's the basic of a functional connectivity. And there's a huge amount of various mathematical and prime mathematics, you know, techniques to extract those networks or those, and, you know, from like a, you know, model-free to like with much more, you know, with many more, like a modeling aspect to it. So if you like, you probably all heard about principal component analysis, you could look at those resting states set of TRs and see whether what are the main source of variation, especially and temporally. That's one way of doing it. It's not the best way. But it's, you know, so all those techniques have been tried, will be tried, you know, on the, to extract, you know, the relevant networks that are actually sharing some information between regions. So let me just quickly go into, you know, a big distinction in this area. So you've got the structural connectivity, which is really like how things are wired using possibly diffusion imaging. You've got the functional connectivity, which is basically, if you see functional connectivity, just replace that by correlation. I mean, that's a good first approximation. You know, you just basically, it's a more fancy word to say correlation somehow. And it's a, you know, it's a bit of a, so how much, you know, those time series are actually, you know, correlated basically. And effective connectivity is a different thing. We are trying, I mean, those techniques are trying to actually have some causal influence. And that's a much more difficult thing than possibly not feasible entirely with FMRI. So it's based on the model. So to just re-establish functional versus effective connectivity, you may have a brain region talking to another brain region which has a correlation between those two things. And A is also correlated with C and is talking directly. There's no direct connection between B and C. However, if you look at the correlation, you will see a correlation because of this, you know. And that's what the, you know, some researchers are trying to decipher, they're trying to sort of, what is the actual, you know, effective connectivity and not the connection in the correlation somehow. So I've got a bunch of things that I wanted to quickly say on functional connectivity. In the previous slide here. Oh, sorry. In the previous slide, how did you get the values of the effective connectivity or? Oh, this is simulation to show you, there are ways of doing that, like for instance, partial correlation is one. I mean, like if you study the precision matrix of a, you know, multivariate normal, the zeros on the precision matrix are, you know, linked to the conditional independence of those regions. So you can, you know, sort of, there are ways of doing that, but that was simulation. Yeah. So quickly, what is, you know, there's a lot of work. I mean, this is not limited to, I mean, this kind of study is not limited to fMRI. People are doing that for EV, GM, and G as well. There's a, you know, popular sort of a techniques. The most popular techniques probably for extracting those networks is the ICS. You may have heard about that, independent component analysis. Correlation in time series can be spurious. I mean, there's a big problem and a lot of literature on the problem of a movement. We say if you move, you're going to correlate things, you know, with, you know, between regions across time, or you could possibly going to decorate things. I mean, that's a big, you know, sort of a methodical hurdle. There's sort of an interaction between, you know, people looking at how much, what's the, okay, what's the correlation during that kind of task versus what's the correlation during that kind of task? So how do I change the correlation, the connectivity, I say correlation now because that's a more fair word, but, you know, between tasks, that's another thing. It's called, it has the fancy name of a psychophysiological interaction. And there's a reason for that. There's the most classic thing is the seed-based analysis. You pick up a brain region, you look at the correlation, and you make a map of the correlation with the other brain regions. That's what is called the seed-based analysis. So I'm going a little bit quickly. Independent analysis, I just wanted to show you the sort of, that is also actually very close to the linear model that we usually perform on that. So you have your data, and again, remember the MEG data, you have the space here and the time here. So this is the big metrics of your data. And then what you're trying to do is decompose that into some temporal aspect and some spatial aspect. So you're trying to decompose that matrix using the independence constraint for that. But there's plenty of, you know, metric decomposition techniques that you could use. I mean, PCA is one of them. And for instance, this is an example of an ICA done on a task-based fMRI where you can see that's two of the components that have been extracted, temporarily show the effect of the task. And especially, of course, as well, this was a visual task and you found this component in the visual areas. But you find other components that are less clear that you wouldn't have found using a specific model because you wouldn't know about it in the brain. So those kind of other components that are unexpected can shed some light on the processing. I'm going to skip that. Just a warning slide maybe on causality. It's extremely hard. I mean, causality, you know, you have to look at what is causality and the definition of causality. But basically, it may be because of, you can believe on causality because of temporal information. You can believe of it because you've got an approri on it. You can believe in it because there's a direct manipulation. If I do this to that node, that is going to happen to the, so, and direct intervention somehow. But in general, effort is very hard to find any actual causal thing with just fMRI. And because of the timing is so sluggish because of all the possible artifacts and so on. So just maybe a warning slide. But some people are trying that. So one of the aspects of causality will be like if something happens before something else, it must be sort of at least, I mean, it could be causal somehow. So people are actually trying to see what's somehow the correlation by shifting the time series and see is there a correlation between shifted time series? That's basically ground causality thing. And you see, I mean, our study is showing that some areas are actually having their activity, again, the bold activity, which may not be the neural activity, but the bold activity before some other areas and therefore trying to find some better model of the processing that is happening in the brain. I'm going to skip that. It's a very interesting technique. Maybe misused, but I'm going to not do that, but that's an attempt to have an actual model that you put the graph. You say, okay, I know that this region is connected to this region. I know that this one is connected to this one and so on. And then I will estimate what is the transfer of information, what is the impact of the, but I'm not going to talk about that. I guess in the five last minutes, so there's, from those connectivity studies and those correlation sort of matrices, there are a lot of people that are actually going from that to graph theory study. And there's a lot of graph, a lot of interest in graph theory studies. And there's a lot of questions. So basically the idea is that you take your brain, you cut it, you parcel it into let's say 200 regions or 1,000 regions or 50 regions or whatever, and then you look at the correlation or maybe partial correlation of this matrix and then you threshold that and you think, okay, this is the graph of the connection for that population, for that condition and so on. And then I'm going to study some characteristic of this graph and there's a lot of people interested in these things. And there's a lot of modern graph estimation. I was talking about the partial correlation, which is related to the inverse governance conditional independence thing. And a lot of techniques for trying to regularize the estimation because you've got a lot of regions. If you do that voxel by voxel you don't have enough information for doing these sort of things. So there's a lot of good interesting statistical methods which is more my world to estimate those things in the best way. It's a difficult problem to estimate the underlying graph when you've got the correlation matrix or the covariance matrix. Many methods don't recover the graph. That's a good study by Steve Smith a few years ago. So it's just to tell you that while there is this attempt there's a small success somehow. And the last thing I will tell you about, maybe not, maybe I'm going to skip that section, but basically I just want to mention that there's a lot of interest in predicting. And it's actually good in terms of science to think that if you have a model that can predict something then it's a much more solid thing that you can work on than to have just something like a map that you look at a correlation aspect if you want. So the prediction aspect is really critical in all science, but in our world, in the new imaging, brand imaging world it's something which is being developed actively for some time. And basically you have several questions about that. You have the question of if the subject, if I have the brand fMRI data of a subject, can I infer what the subject is doing? So it's kind of the light detector or the machine, can I know what you're thinking when I have got your... And obviously this is fantasy. You will never know exactly what the subject is thinking of. It's not going to work. But there are some things that you know that there's a pattern of activity that resembles that you're pretty sure that the subject is more looking at a face than a car, for instance. That kind of thing can be predicted somehow. So in control environments, you can infer some information with fMRI. And that's what... So I'm skipping that. There was an early study by Jim Hacksby on that program, like looking how much the pattern was looked like, a pattern that is... Like the pattern that is in the physical face area when you're looking at a face, for instance. And therefore being able to say, okay, the subject is actually now looking at a face. And there's an interesting work in this area where people would... Okay. So the last thing, which I'm now over time, I guess, but is reproducibility issues. So there is a lot of... I mean, fMRI and brain imaging is a very sort of popular technique. It's always very easy to put a picture of the brain with a shiny spot on the anatomy in the newspaper. So therefore it has a lot of a hype somehow. But the reality is that because of that partly, the number of studies are actually underpowered. So you find something by analyzing a lot of the data, and then you eventually find something. And because you're a PhD student and you really have to publish, you're publishing something, but there's a lot of unreproducible or difficult to reproduce or uncertain results in the literature. So the last thing I wanted to mention is that there's a very strong feeling in many of us in the community that we should improve the brain imaging reproducibility aspect by a number of things, like having bigger ends, bigger group population, sharing the data and the code so that people can see what's being done and possibly find mistakes. It's so easy to make mistakes. It's a complex set of number of processing that you have to deal with and it's just very easy to make mistakes. So having a better sort of ethic on the data sharing and the code sharing and thinking, okay, this is, as a scientist, I should be making sure that my results are valid and are reproducible. And therefore, I mean, I've got an interest in having the feedback of the community on those things. So yeah, that was the last thing I wanted to mention that brain imaging has probably, along with other scientific communities, to work on the reproducibility aspect of some of the studies. And it's a general thing, but the specificity in brain imaging is that we've got usually a small group of subjects and therefore a very underpowered study. So that's computing the power of your studies is one of the critical things. Let me feel we'll start. And there are some good initiatives like there are big data set that are being required and shared. And so the world is a bit changing. And I think there's an emerging somehow skill and there's an emerging population of scientists who a little bit like in the genetic community, we've got biostatisticians that are, they know the databases, they know how to query those databases, they know what's interesting in terms of the all this information spread across all those resources on the web. I think the brain imaging is actually moving towards this. We're going to get like informaticians working on a system to share, to query databases, to have all those ontology that is getting there. And Marianne, I don't know if you've talked already about that, but you will. And so there's an emerging sort of a scientist, a kind of scientist that will have a hybrid sort of a skill and actually both skills of knowing enough on the brain and the neuroscience aspect and knowing and having good skills in terms of a database processing, computer science and the knowledge of where are the resources and how can I use that to answer good questions about the brain. I think that's an emerging sort of a job somehow. That's a wish some of you will be interested in. And I think I will finish here. Thank you for your attention. Which one? A different activity if you see this house or anything like there's on slide. Brandy coding, the XP one, yeah, this one? Yeah, yeah. So do you mean to say like if the person is seeing face or house, there is a different pattern of activity? Right. So what if that person is recollecting those same images? Do you see any difference in the activity? Yeah. Like seeing a house and recollecting the same house in my mental imagination? Yeah, I think, I think I'm not 100% sure but I believe like just recalling and just mental imaginary of a specific face versus mental imaginary of a house will probably lead to some success in the coding. It's not going to be 100% but I think there is, yes, I think there is probably enough activity in the fusiform face area compared to the peripocompole place area, for instance, that you may be able to be able to say, okay, this subject has more likely thought about the house versus the face. That's probably correct, yeah. But I should, we should look at the literature for that but that's my take on it. And also there could be associations. So this depends on person to person, right? You cannot generalize. So for example, if I see the nest of a bird, maybe I could have the same activity because I have a very association of house and the nest of a bird or it could be something like I see a house. So how, yeah. So now again, what I was saying is that in very controlled sort of environment where you know that the subject is either thinking of a place or either thinking of a face, then you may have some success, you will have some success in decoding, in a sort of a decoding is bad work but predicting what the subject is actually thinking of. But in general, I mean, in general, that's just a task that we are not able to do and it's going to be, yeah, it's... But still it's very difficult to have that controlled environment because in yesterday's lecture, I think in one of the example, you're showing like a person was shown an unknown word. When you say unknown word, it could be possible that he has forgotten that word. Maybe later he realized how you will ensure that it's unknown. So maybe I saw a face today. It's totally new face for me. But after some time when I started talking, I realized, oh, we had met 20 years back or something like you recollect. Again, again, it's under very controlled, specific experimental design, you know. It's not, I'm not sure what that answer is. No, I understand that, but how do you ensure that it's very controlled? For example, how do you ensure that that word is unknown to that person? So that's the word of an experimental psychologist and they are skilled with that and they have some, they do sometimes other experiments after the scanning or before the scanning to get those control. It's a whole sort of trade to sort of making sure that what you think you are studying is actually the thing you're studying. But there are, I mean, I'm not saying in general, it's a... So you mean like it's possible to have a controlled environment precisely or I just want to know how much percent is there any deviations? I don't think it's not possible to control what the subject is thinking. No, means to know exactly. It's not possible to say for sure that, you know, what I'm saying is that I think it is possible to ask subject to think of either one thing or the other and to some degree recover which stimulus the subject was thinking of. That's all I, that's the only thing I'm saying. I'm not saying, you know, like in other conditions, in like it's a... But, and again, I mean, how much of that you can recover? We should look at literature, maybe a small amount of that. I mean, it's not a machine that reads your thought contrary to, you know, newspaper, sort of. On the topic of reproducibility and unworthy publications, I think it's especially present in fMRI studies because it's typically very low subjects and high-dimensional problems. If you know the ignoble prize, like an anti-noble, there was a lot of publications from fMRI studies that won the prize. And it's a fact that it's due or so... I mean, you've read the UNID's papers, but if you haven't, do read the UNID's 2005 and, you know, and why the most research results are false and it's a combination of things. And I actually have a few slides on that. Actually, maybe I should show you at least the... So, yeah, that's a... And it's a combination of things. And one aspect is a sociological aspect that, you know, many people are, you know, getting like a reward of publication and things. And they're trying, you know, experiments without much thought or like much seriousness. And, you know, one aspect of that is that, you know, there's going to be... The other aspect is like the flexibility of analysis. I mean, when you've got the pressure of the scientific community to publish, because you have to do that if you want to stay in academia and, you know, then, you know, you will publish something eventually by squeezing the data as much as you can, and, you know. And paper journals don't take very much into account, like, non-positive results. And, you know, there's a whole bunch of reasons and rational why there's a problem. And some scientists would actually partly deny that because there's a very strong feeling that what you're doing is right. And, you know, it's very hard to sort of have the humility to say, okay, you know, I should really focus on the facts and on the... It's just human nature. I mean, it's hard. But as a scientist, I think that's the... You know, you have to think of, you know, okay, I just have to be, you know, like... And it's difficult. And so there are solutions that are emerging. Like, so there's a number of people who say, okay, we should really share more of the data and the processing and things like that, so people can check. There's also the pre-registration of hypotheses. For instance, if I have an hypothesis, I should pre-register it, and then I'll cry the data, analyze the data, and say, okay, that was wrong, or that was true. And like, you know, I confirm or not the hypothesis. That's a better sort of a path. And some journals are going this direction and they say, okay, you know, if your hypothesis and the method that you're going to investigate it is okay, that is reviewed by some of your peers, and if they agree that's a good hypothesis, that's a good method to investigate it, then whatever the results, we're going to publish it, even if it's a negative result. And so there are ways of... Yeah, well, some journals set up publishing, like a published checklist, like Nature, for example, so it's really controlled the statistical analysis, but it should really become a trend in more journals, I think, and they should also employ the statisticians with it. And just a quick question I want to ask in the... I'm excited you're getting worried about the time. Are we okay with time, or who's in charge of a... Who's the timekeeper? Who's the timekeeper? Okay, okay, but... Okay, well, just about the anatomical atlas, so I guess the sub-cortical areas were probably estimated anatomically, so edge detection, image processing. What about the cortical areas? Is it more functionally determined? It depends. It's a good question. There are some deep surchar that are actually good landmarks for change in functional areas, but it's not necessarily always the case. So it's a bit of a complex, mixed... And it is a research question, so, you know, it really is a research question, which... We can ask you later, one else. I also say one thing about not producing full results, and at least from my experience, I've never done this kind of things, but if you do things without completely understanding it, I mean, borrowing some methods from others, some library or something, you're much more likely to do mistakes. So I think this is also that this is sort of like this very advanced machine, and a lot of the people don't... Haven't really looked into the basic physics of it, what you really measure. So just do this as a black box, and I think that's also a recipe for making mistakes. It is true. And to, I mean, making mistakes or making interpretation mistakes, yeah. But a lot of mistakes are done during the processing. I mean, let's assume you know roughly what you're measuring and you're not going to over-interpret the bold response. There's a lot of room for mistakes and the processing levels. And most of the psychologists or even the MDs that are working with those tools, they often lack the appropriate training on the computer science aspect that would sort of limit those mistakes. So we need these people, right? We need these people, absolutely. Coffee, I think? Coffee, okay, you're the antique. Okay, thank you.