 Pleasure to meet everyone. My name is Jonathan Uriartalopes, but feel free to call me John. He's what I go by. I'm an undergraduate researcher at the Developmental Cognition and Neuroimaging Lab, also known as the Deacon Lab over here at Oregon Health and Sciences University. Just up the hill here in Portland. I'm an undergraduate researcher up there, and my my project is going to be about improving the quality of neuroimaging scans. So without further ado, let us begin. So here's a little bit of an overview of my project. Our main project is to study the brain, specifically subjects who have been diagnosed with autism. Now the problem with autism is neurodevelopmental disorder, so we need to study the brain, but we're humans, we're not cars, we can't just like pop the hood, see what's going on with the engine, then pop it back up. So that's why we use this technique called MRI imaging, which stands for Magnetic Resonance Imaging, and this is done so we are able to take images of the brain safely and in subjects. So ASD, also known as autism spectrum disorder, as I mentioned, is a neurodevelopmental disorder. It's something that the patient has and is stuck with them for the rest of their lives. And it's commonly characterized by these following symptoms. They have a hypersensitivity to stimuli, whether it be light, sound, a lot of people around. They're just able to feel them a little bit more. And they also have a restricted repetitive behavior, so they tend to follow the same routine over and over again. And in some cases, they also suffer from social and or a language school deficit, which makes them hard to be able to communicate with their peers and engage in the environment. And there are many reports about the approximate diagnosis rate in the United States, but I was able to find that affects approximately one in 59 children here in the United States. So it's a fairly common neurodevelopmental disorder within the community. So because, like I mentioned, it's a brain disorder, so we want to try to study it, which is why we use MRI imaging, because we are able to do it safely. And by able to use MRI imaging, we are able to get subjects who have been diagnosed with ASD, compare them to brains who are typically developing, also known as TD subjects, and try to see what are the difference between the two. Unfortunately, MRI scans are expensive and technical. They're very big, clunky machines. The latest three Tesla machine is about $3 million, and in case you guys didn't know, it's about $500 per scan per hour, so it's really, really, really expensive. So because of this, a lot of neuroimaging papers have difficulties getting with data acquisition, and low sample size is a common issue. So in order to solve this low data acquisition issue, we turn to this international neuroimaging data set known as a BYD. A BYD stands for Autism Brain Imaging Data Exchange. It's an international data exchange program created by Dr. Adriana Di Martino and Dr. Stuart Motowsky to try to solve this small data issue acquisition. It has 25 international institutions and there's more than 2,000 MRI subjects, subjects who have been diagnosed with autism spectrum disorder, as well as subjects who are just typically developing normal kids. And these neuroimaging scans contain structural scans which allows you to be able to see the 3D volume of the brain, and functional scans which allows you to be able to measure brain activity within the brain. So like if you tap your fingers, your brain signals pop up and we're able to detect them. This is a great data sharing initiative because some scientists tend to be a little bit stingy with the data they don't like to share, but because of this large reservoir of data, neuroscientists around the world are able to use this data to be able to conduct their own research and this is great because by having a lot of research to look into this data, we're able to see distinct differences between autistic and typically developing subjects. But before we can use this data for analysis, we had to go through a quality assessment protocol developed by the lab. This quality assessment protocol is called the Standard Operating Procedure for Quality Assessment. It was created by Emma Schiffsky and the rest of the community team at the DINCA lab and the purpose is to quality control also known as QC neuroimaging data. This is used to analyze the abide data sets because we want to make sure that the data set is usable for analysis and but before you can become research reliable you must undergo a training, you must take a training set, pass it, you got it, you're given a list of subjects, you got to have an 80% success rate before you can be considered certified. So why do we want to quality control this data set? They give it like cooking, you're trying to make a meal but before you can make a meal, you want to make sure your ingredients are good. You got good vegetables, good meat and all that otherwise if you try to use like nasty raw meat when trying to make a hamburger, you're going to get sick. So we don't want none of that. So the QC outline follows up three main, I guess you can say subject. There's the Atlas Registration where basically we get the raw T1 metadata and we match it to a template known as the Montreal Neurological Institute also known as MNI template and we try to see if they can match up. Kind of like making sure your left hand matches up with your right hand, make sure your right hand matches up with the left hand, make sure that the maps are all on the same page. The second part of this QC outline is the structure data where we look at the 3D volume of the brain and we check out the delineation of the gray and white matter. Those are the two types of brain matter that's within the brain and we want to make sure that they're delineated or mapped and bordered correctly. That way we can try to see structural brain differences between the two types of subjects. And the last one is functional data which I mentioned previously is to measure brain activity. So what's going on with brain activity and ASD and TD subjects? That's what we're trying to measure. Try to see if there's any other reasons that might be able to explain their symptoms. And this QC scoring outline is, guideline is really, really simple. You got three numbers. You got one usable. It's like getting an A in your class. You got two which is probable. It's like barely passing with a C minus. It may or may not happen to me. And then you have three which is unusable. It's completely not able to be used for analysis. It's kind of like getting an F in the class. So I'm going to give you guys three examples of a usable, probable and unusable scan. So I know this is a pretty lot of information but right now I want you guys to focus on the unusable scan which is the scan on your right, my left. So if you look at the frontal lobe which is like the front part of your brain, right by the eyes on the left, left this portion of the brain, you can see that the red lines, it does not really capture a lot of the gray matter. And that's bad. It's like trying to catch a bunch of fish but you miss most of them. And because of that, that scan is unusable for analysis. And then if you look in the back portion of the brain, you can see it's not straight like a brain is supposed to look up like the skull is supposed to look like. And this could be due to because of warping. So when you're inside the scanner, you're not supposed to move. You're only allowed to move like 200 micrometers which is like this. And the reason why you don't want to move in the scanner is because the MRI machine will pick up a signal here. But if you move, it might think it's right here. So that will cause warping where it's just going to stretch out. And you don't want that in a scan and that's why this scan was great as unusable because it will be the possible movement within the scanner or maybe processing pipeline issues. And then the image right underneath it is an example of a functional scan. This is really, really bad because you're missing about half of the brain. And if you don't believe that that's a bad scan, you try to do this presentation with only half your brain power. So in the second scan, you see the probable which is like dead straight in the middle. If you look at the green circle on that image, you can see that in the frontal lobe, the delineation between the gray white matter is good. But the reason why this is a probable scan is because if you look at that yellow circle in the middle, the delineation goes into the dura and that's also a possible sign of maybe moving within the scanner which disqualifies this from being a usable scan. And then if you look at the functional scan on the bottom, you can see that there are some areas where the red lines does not fully encompass the brain whether it be the superior parietto or the frontal lobe. And because of that, it's not quite usable for analysis. Then the usable scan, it's clear, cut, simple, easy going. Everything is what it's supposed to look like, so that gives it a one. And the functional scan on the bottom, there's a slight issue with the left temporal lobe but that's perfect for analysis. So next thing I'm going to talk about is our decan computing pipeline which we're going to be using to try to improve the MRI scans. So there are going to be two main approaches. Approach one uses the original human connectome project which is the ACP minimal preprocessing pipeline. This pipeline was created by Glasser and his colleagues to try to process MRI imaging scans. This is the latest state of the art pipeline and because of this, a lot of individuals are able to use this pipeline to create high quality imaging scans. But the issue with this by data set is that the scans used for the scans when they were acquired weren't originally made to be put through this pipeline. And because of this, you might get some issues that I will explain later. And it's kind of the scans were obtained in the 90s and they were used with older MRI scanners and this pipeline was designed to be able to process data from newer up to date scanners. So because of that, you may have some issues which is why it's kind of like trying to put a snapshot on your old flip phone. It's not going to work. So our approach two is going to be using our modified pipeline. So we grab the same ACP pipeline but we're going to tweak it a little bit. It's like grabbing a recipe for a dish but we're just going to adjust the recipe and try to see if we can make it taste better. So we're going to use that same ACP pipeline but we're going to add this algorithm known as advanced normalization tools also known as ants. This is an algorithm created by advance and his colleagues and we're going to use that algorithm to try to see if there's any changes within MRI quality imaging and try to see if there's any way to see any notable differences and this is going to be our decan pipeline. So this is going to be our pipeline overview right here. So I'm just going to go over this briefly and if anyone wants to discuss this pipeline further feel free to let me know during our question seminar. So we have the pre-free server pipelines basically where we grab our raw DICOM files that we get from the MRI scans and we generate the 3D volume of the brain and we match it to a template like the Atlas registration I mentioned previously. And then you have the free server pipeline which is right here in the middle. This is where you grab the brain volume and you delineate the predefined structures. This is where we get the gray and white matter going on in the brain as I mentioned in the structural scans. And the last portion is the post-free server pipeline and that's where we grab all the MRI data and we convert them to the CIFTI files which is used to store the data. And the second part of the pipeline is the HCP functional pipeline and this is where we're going to be able to get all of our functional data. So we've got the fMRI volume data where we grab the, this is where the part of the pipeline where you grab all the fMRI data and you normalize it across the 4D structure of the frame. Now what do I mean 540? You grab the 3D structure of the brain with respect to time and we try to use that to measure brain activity during tasks while the subjects at rest and in response to stimuli to try to see if there's what's going on with the brain activity and how the brain's working when they're engaged in possible environmental stimuli. So this pipeline that you see right here is approach 1. You see your original HCP pipeline and that's what Glaston and his colleagues were able to do that. And I'm going to mention the two ANTS corrections that we did to try to improve this pipeline. So ANTS corrections 1 is going to occur right before the pre-free server. The purpose of this algorithm is to denoise, denoise the data, basically try to get, so denoising is where we grab, is where we try to increase the signal to noise ratio. So we try to make the scans clearer, more easy to see, and make it not so ringing so you get able to properly delineate the gray, the gray and white matters throughout the brain. And then the bias field correction, it's where we try to, it's where we try to grab, grab, the brain's supposed to be like right here and if it's a little bit out, it's kind of like going to the chiropractor, getting your spine realigned and just pop it back into place. That way we can make sure that the portions of the brain that are supposed to be there will be there. And the second portion of the scan is to use ANTS correction 2. The purpose of this, this portion of the algorithm is to adjust the scan better to the atlas. Make sure, again, make sure that aligns better because if it aligns better, we make sure that all the parts of the brain that are supposed to be there are going to be there, which is going to end result, going to give us more accurate, higher quality scans. Now I know I went through this and mentioned through a complicated pipeline, so what does this mean for the audience? So we basically grabbed this raw DICOM file that you would get from the MRI imaging. We ran through this pipeline. You're able to get the structural scans from the first portion of the pipeline on the top and also generates the functional scans from the HCP functional pipelines on the bottom. After implementing this pipeline, reprocessing the scans, my next step was to try to figure out, alright, how do they compare? How does this do these ANTS correction? Do they make the MRI quality scans better? Do they make them worse? Do they make them the same? And from what are you able to find with a couple of subjects that it actually makes them a little bit better? The brain image on the very left is an example of a scan that was processed with the original HCP processing. Now this scan is now horrible, but as you can see, it's not quite clear. It's pretty ringy. The red lines miss some of the gray matter, and it's not quite as clear. And if you can see the circle on the red circle, you can see that the skull shape is not very straight. It's not very curvy. It's kind of wavy. So that can either be, again, either through moving with a scanner or because of processing pipeline issues. So in the brain image on the right, that's an example of a brain scan from the same subject, but instead it was very processed through the new decan pipeline, the HCP pipeline with the ANTS correction. With this, you can see that the gray matter, the brain scan overall looks a lot clearer. And if you look at the green circle, instead of having that wavy portion of the skull going like this, it's nice and curved like it's supposed to be. And then if you look in the blue circle, you can see that a lot more gray and white matter is captured, and the scan looks a lot clearer. So after noticing this with what the bunch of subjects, I did a quality control analysis using the same protocol I mentioned previously of an abide data set that was processed originally with the HCP pipeline and compared it to the new pipeline that we use right here. So I did a quality control of about 180 subjects, and I did that twice. So I looked at roughly 360 scans, tried to see like, all right, how does it improve? And what was really, really interesting is that this improved it greatly. So if you look at the pink color or salmon colored portion of the graphs, those are the part that's the numbers used to represent the modified pipeline, the new HCP pipeline with the ants corrections. And the blue cyan colored are the original HCP minimal preprocessing pipeline. And I'm not sure if you can see it very well, but they were graded with usable, probable and unusable. They're labeled on the bottom. If you were to see the unusable rate for both pipelines, you can see the unusable rate is 36% for the original HCP pipeline. That's more than a third of the data set. But the unusable rate for the new decan pipeline is only 7%. And then if you look at the usable rate, it increased from 36% to 53%. So it was really, really, it was a drastic improvement. And to everyone in my lab, everyone was happy about it because we said, hey, we did good science, everything went well. And then I know there's going to be statisticians over there, and this is science, we need to use stats for everything. So around the chi-square analysis, and we were able to get a p-value of 3.5 times 10 to the negative 8. So in other words, it's really, really statistically significant. So this is great. That means we proved that this pipeline improved. We showed that there are data issues with the abide data set. But that doesn't mean that this data set, all the MRI images there, is unusable for analysis. We're able to grab that data set, refine a little bit, and then we can try to improve the data quality and then possibly release it back to the public. But instead of having these iffy scans, you're going to have a more reliable and more accurate scans to be able to give to the community. So just to give a little recap, near imaging is a very, very important technique. It's great. MRI scans, we're able to see the brain without having to cut it open. But as I mentioned previously, MRI scans are very, very expensive. They're very, very technical. They're hard to run. And then also trying to process the scans, take a lot of energy, a lot of manpower, and it's very, very tricky. It's very, very sensitive, very, very delicate. So you need to be careful with it as well. So the abide data set is a great publicly available initiative because, as I mentioned, that's about 2,000 subjects. So if you do the math, every scan is about, let's say, 30 minutes. So that's about 250 times. Let's just say 2,000. Actually, I'll take the back. I'm not going to do the math. We save a lot of money and a lot of time. I would use a calculator but I don't got my phone on me. But yeah, it's a lot of scan data that we're able to do. And then also it's really hard to get subjects that are who are diagnosed with autism and also typically developing because you got to go through this whole process, this whole protocol to get them into the universe, to get into the scanner. And not only that, you need to make sure that the children are not moving in the scanner. Otherwise, it's going to affect data quality. And so, and that's one of the reasons why I wanted to make this move pipeline because I'm not sure if you guys remember when you were kids. It's very, very hard for us to sit still. Very, very hard for us to do what many of you guys are doing very greatly. Just sitting still, listening, paying attention. Because I remember when I was a little kid, I liked to move around a lot or I could distract about a little thing and just go left and right. But because of that movement, we're going to have a data issues when we collect these scans. So the improvements to this pipeline is still ongoing because as I was just showing right here, we had very, very good improvements, but that doesn't, but we're not going to stop there. We want to get better. We want to try to see what can we do to improve that because these issues are not going to go away anytime soon for MRI scans. So if we were to figure out like little tricks, little kinks, little things that we can do to improve data quality, we're going to do them. We're going to do that. And not only that, but also share them with the community so they're like, hey, this is what we notice with our scans. Maybe if you do this for your scans, they might improve as well. And that way we are able to go back to the main question of like what's the difference between brain development or brain structure of subjects who have been diagnosed with autism spectrum disorder and subjects who are typically developing. So what's causing their symptoms? What's going on right there? And a lot of scientists will like this new pipeline because it will save them time, money, resources, and most importantly, data because that's the big deal out of all this. So this is the end of my talk. But before I take questions, I want to give a special thanks to Oregon Health and Science University for providing me a great workplace and also the Deacon Lab. I had way too many individuals who have been patient with my many, many questions. I'm a biochemistry major by practice. So going into this world of computer science, MRI scans, and I'm going to be honest, I wasn't that great in physics at school. So I know how many great people on the community helped me out. I had my mentor, Dr. Eric Fesco helped me out, Eric Earl, Derek Sturgeon, Emma Schiffke, Anders Perron, and many, many, many other people I can mention. And they're all shown in this picture that they've been able to help me out. And as I mentioned, I'm still a student and I'm not sure if you guys remember, college is expensive. So I want to give a special thanks to the Robert E. McNair Scholars Program and also the Billy Zeta Program for offering me scholarship money and a research opportunity to do my incredible research that I'm doing up there on OSSU. With that further ado, I'm now open for questions. Thank you for your time.