Hi. In this module, we'll talk about the analysis of fMRI data. We'll talk a little bit about some nomenclature and we'll also talk about the goals of fMRI data analysis. So functional magnetic resonance imaging or fMRI is a non-invasive technique for studying brain activity. By non-invasive, we mean that there's no known side effects for taking frequent fMRI scans. So during the course of an fMRI experiment, a series of brain images are acquired while the subject performs a set of scans. Then changes in the measure signal between the individual images are used to make inference regarding task-related activations in the brain. So as a simple example, let's say that you're doing a finger tapping experiment, where you do finger tapping for 20 seconds, followed by rest for 20 seconds, followed by finger tapping for 20 seconds. Under this setting, you might measure the brain multiple times while you're tapping and while you are at rest. And then you look for differences in the measured signal between activation and rest states. So fMRI is again, functional in the sense that is measured continuously over time, so you measure the same brain volume multiple times across time. Each of these brain volumes consists of roughly 100,000 different voxels which are cubic volumes that span the three-dimensional space of the brain. So consider that you have your brain and you put it inside sort of a bounding box. And then you take this box and you split it up into 100,000 equally sized boxes. Each of these are the fundamental unit that we work with in fMRI. They're called voxels or volume elements. Each voxel has a number associated with it, but also a spatial location. So basically, in this little cartoon, we take away a voxel from a certain area of the brain and we look at its value. In this cartoon, it's 39. So again, each voxel corresponds to a spatial location and has a number associated with it that represents its intensity. During the course of an experiment, several hundred of these types of brain volumes are acquired, one roughly every two seconds or so. So basically what we have is we make 100,000 measurements over the brain at one time point. Then two seconds later we do it again, etc., etc., for a couple of hundred time points. So another way of looking at this, is we can extract the information from a single voxel. And as I said, a single voxel represents a spatial location. So if we take the same voxel across time, we're actually studying what's going on and how the intensity is changing across that voxel in that spatial location. So by doing this, we can extract the time series of these intensities and study to see whether or not there's something in that time series that's related to the task that we performed. So in my little example, I was saying we were doing finger tapping, resting, finger tapping, resting. Then we might look for a voxel where the activation is going up while we're finger tapping and going down while we're resting. Such as in this little cartoon where we see this sort of boxcar activation. So one of the interesting things is that, this shows you that fMRI data analysis is fundamentally a time series problem. Because the data from every voxel is a time series, in this case. However, it's sort of a time series problem on steroids. Because what we have is, every voxel of the brain has its own time series and there's about 100,000 different voxels. So basically, we're dealing with about 100,000 different time series that we're studying and looking for at a task-related behavior. So what is this signal that we get in this time series mean? Well the most common approach towards fMRI I used is what's called the Blood Oxygenation Level Dependent or BOLD contrast. BOLD fMRI measures the ratio of oxygenated to deoxygenated hemoglobin in the blood. It's important to note that BOLD fMRI doesn't measure neuronal activation directly. Instead what it does, is it measures the metabolic demands or the oxygen consumption of active neurons. Where neurons are active, they need access to oxygen to replenish their energy. And it's this oxygen consumption that we can see, so which is a side effect of the neuronal activation that we're actually interested in studying. So basically, the way the signal changes in reaction to some task is described as something called the hemodynamic response function or the HRF. This represents changes in the fMRI signal that's triggered by neuronal activation. So let's say that I do something like clap my hands very quickly. Then neurons in my motor cortex start firing and this leads to an increase access to oxygenation in that region of the brain. So basically what we'll see, is we'll see a signal changing in a manner that corresponding to this cartoon image here. We'll see a rise in oxygenation levels that peaks after five to six seconds which goes down. And after about 10 to 15 seconds, it goes below baseline and then returns to its baseline form after 25 seconds or so. So clearly, fMRI data analysis is a massive data problem. Each brain volume that we're studying consists of roughly 100,000 different voxel measurements. Each experiment might consist of a 100 of brain volumes. And each experiment might be repeated for multiple subjects, maybe say 10, 20, 30 or 40 subjects. In order to facilitate population inference. At the end of the day, the total amount of data that needs to be analyzed is staggering. The statistical analysis of fMRI data is challenging. It's a massive data problem. And so it's a sort of like one of these modern big data problems that's facing Statistics today. Also, the signal of interest is relatively weak and the data exhibits a complicated temporal and spatial noise structure. So throughout this class, we're going to start trying to discuss these different things and try to understand how we can analyze fMRI data using statistics. Throughout, we're going to keep coming back to this data processing pipeline which illustrates the different steps of an fMRI data experiment. And these consist of the Data Acquisition and Reconstruction step. There's also a fair amount of preprocess into the data undergoes. And here, we'll talk about things like Slice-time Correction, Motion Correction, Co-registration, Normalization and Smoothing. And then finally, we'll talk about Data Analysis and there's going to be a couple of key goals that we'll talk about shortly. Throughout, experiment of the design plays an important role. So these are all different types of things that we're going to be covering in future modules. >> There are three main goals of fMRI data analysis. The first one is Localization. The process of determining which specific regions of the brain are active during a specific task or in relation to a specific psychological event or behavior. And there are varieties of localization. And all of these together constitute what we call the Brain Mapping approach. So the basic framework. Let's start with the assumption that we have some signal on the brain and some areas that are truly activated. Those are shown here in blue. Those signal areas, when we conduct scans on real people, are mixed together with noise which is smooth in time and space. So you see here the red noise pattern. Where bigger circles mean higher, more positive noise. What we observe is a mix of the signal and the noise together. When we do a statistical test, we usually do it at each voxel. It's called the mass univariate analysis. And we'll deal with that more later. And we essentially will conduct a hypothesis test at each voxel. Because we've made many comparisons, we have to correct for those multiple comparisons. And so we end up observing, usually a small fraction of the areas that are truly active. And hopefully, in this diagram, all the areas that we've picked out as results show some true signal. That is there's some blue in there. But as you can see, there's one or two areas here that are false positives, so there's no signal at all. This brain mapping approach applies whether one is looking at results from a task. This might be a task minus control comparison in which each point here in one voxel constitutes a set of scores from individual subjects. And we're doing an analysis across those subjects. That also applies to a Brain-Behavior correlation. Let's say I have a number of subjects again and I'm correlating brain activity with some behavioral measure. It could be their age. It could be their performance on a cognitive test. And finally, it applies to what's now called Information-based mapping. Which is the idea of taking a local area of the brain and then figuring out how much predictive accuracy that brain area has for a particular task, condition, or contrast or behavior. And we'll do which called the searchlight in this approach. And move a window around the brain and make a map at each voxel of how much local information there is. In all three of these cases, we're essentially doing the same kind of brain mapping procedure. And the same principles apply. A second thing we can look at is connectivity. fMRI in particular, gives us the ability to measure or assess how brain visions are functionally related to one another. In connectivity here, usually means correlations across measured values, across time. As opposed to functional connectivity in animal studies, where they're studying direct neuro connections from point to point. And there are many varieties of connectivity that we can assess. Here are the three popular ones. The first one were called Functional connectivity which relates to correlations across time. There are varieties of this. One is seed-based connectivity, in which we take a region of the brain and we ask what else in the brain correlates with that region. A second kind of connectivity is called Effective connectivity and this subsumes many families of statistical models. Including Path analysis and mediation models, Granger causality models and Dynamic causal modeling, to name a few. We'll learn about each of these things later. Most of them in part two in the next course. And finally, there are models of Multivariate connectivity. They fall into two types of categories. One is the data reduction type, which includes techniques like principle components analysis, independent components analysis, versions of those. And the second is Graphical models, which refers to the process of constructing a visualization and it announces framework consisting of nodes and edges among a number of regions. And then deriving properties of that. That describe and potentially have some inferential power for explaining behavior. Finally, the third type is for Prediction. We can use a person's brain activity to predict their perceptions, behavior, health status and a number of different kinds of outcomes. So on the left here you see an example from our work. Where we've developed the Classifier Pattern to predict how much pain is somebody is feeling in response to a given stimulus. And we can apply that pattern to new brain images coming in to make a prediction about how much pain they're feeling. And we can validate that prediction across individuals, across studies and so forth. There are many emerging applications of prediction that cover really the whole space of different possible things we might care about in terms of health outcomes. So this is really an exciting area in terms of translation and clinical science. So some of the emerging applications are to Alzheimer's disease, to depression, to chronic pain and anxiety, to neurological disorders like Parkinson's disease, to behavioral issues like substance use and abuse and to basic emotions and other areas of research. So that's the end of this module. What we're trying to do is give you an introduction to fMRI data analysis and what some of the major goals are.