 The last presentation is Rami Gabriel who comes from I come from UC Irvine It's actually a COA prior to medical school. Is that correct? So he's in my clinic refracting patients I was shocked and then I realized he had had all this prior ophthalmology experience So we're grateful to have him come rotate us with us on this month on rotation and turn it over to Rami Thank you, Dr. Jarnian So I'll be presenting on machine learning and it's applicability in OCT and geography I know Becca Gentry kind of gave a topic similar a while back. So hopefully I won't bore you guys too much So the objectives understand a little bit about machine learning Learn about what Google brain is doing an ophthalmology Learn about OCT and geography and how it works and then discuss some of my research. Hopefully you can use some of these tools in your own research as well So what is learning besides being avoided by medical students? so machine learning is basically a way of acquiring knowledge and you can do it through experience Or you can do it through study machine learning is solely through experience So there are three main categories of machine learning you have unsupervised learning And so that's where you give it data without any answer to the data So say you give it farm animals and then it groups it by those farm animals However, it kind of finds fit so it could group it by color instead of the animal itself like horse sheep And then you have supervised learning and so that's where you give it data that has an answer And so it learns from that data and then it learns to choose it better You have two different types now. We're going to be focusing on supervised learning for the majority There's classification and so that's different categories and there's regression and that's when you Have more of a linear kind of data. So age cost Things like that. The bottom one is actually how I first got involved in machine learning. It's reinforcement learning That's mainly for games and more real-world kind of applications. It has a high-level Goal and so something like chess and so alpha zero from Google beat stock fish, which is the engine for chess Which was amazing So what is Google doing an ophthalmology? The first these are two of the major studies the first one on the top left top right top left for you guys is Basically looking at 128,000 fun disc photos And assessing diabetic retinopathy and actually did a really good job at whether it should be referred or not referred The sensitivity and specificity Are both up in the you know high-middle 90s And actually it did a really good job Predicting the stage of diabetic retinopathy as well. And this is actually FDA approved as a referral for sending to retina specialists The second study Borrowing from Catherine should really be called fundus photos more than meets the eye So it looked at fundus photos and it assessed things that really we didn't expect were To be able to be learned from fundus photos. So one of those is age Gender blood pressure and so it did a really good job predicting age between An age range of five years had 78% accuracy But as ophthalmologists you guys know fundus photos is really just scratching the surface No pun intended. So we have OCT where we can go into the layers of the retinal and specifically we chose OCT angiography OCT angiography is similar to other spectrum domain OCTs It runs the a scans lot faster. The Heidelberg runs around 40,000 and the CRS is around 27,000 And how it works is basically It sends multiple Scans and light and then the amplitude coming back is what's measured For the angiography the SSAD a which is split spectrum amplitude Decorrelation algorithm basically looks at the change in the amplitude. So Tissue that doesn't move should have a constant amplitude But tissue that's there then not there will have an amplitude that varies and so the decorrelation is basically one minus correlation So if the amplitude is all the same you won't get much of a signal for things moving But if it's very buried then you will get something that moves and so you can see on the far right on the bottom image You can see some of the superficial vessels and you can see the Corio capillaris And that OCT is amazing. You can get an in vivo histology And so that we are looking basically at two capillary plexus in this data the superficial capillary plexus and the deep capillary plexus I'll look at it in the nerve fiber layer and the internuclear layer respectively And this is what it looks like At its end product the superficial capillary plexus at the top and then the deep capillary plexus at the bottom right This is actually from our lupus study And so the lupus is on the right and then normal is on the left and you can see some kind of rare refaction in the vessels All right, so our OCT a data we had about 1800 scans. We had About 900 patients and the mean age was 70 But we had quite a quite a large age range from 8 to 99 And then the majority of our patients were macular degeneration. They're all diseased in the In the prior studies with google. Most of them also had diabetic retinopathy So this is just a correlation heat map. You can see if something's highly correlated Then you will get A lighter color and so these are L1 and l2 are just superficial and deep capillary plexus respectively and then the whole image or the Or the hemisphere and separated by different sections of the OCT a image This is a box plot. You can see age with age over 70 and under 70 You can see somewhat of a downtrend These are all just box plots, so it's not significance here and then gender Kind of similar but not as big of a difference So let's apply our machine learning. What exactly is it? So we used a multi-layer Perceptron algorithm and so basically what this is is The thing on the far left those x values are inputs and so we had 19 the Different quantitative values from the OCT a and then the second is the layer And so you basically assign a weight to all those inputs and then you assign Those weights into the middle node and so you kind of create this network And then those weights then go on to the final output And so that's where kind of neural network came in is that each one of these is like a neuron for predicting gender We did a terrible job. We got 53 accuracy So we kind of gave up gender pretty quickly and then for age we had 74 accuracy, which was actually pretty good This is a confusion matrix. This is actually what it's called But don't be afraid It's actually very similar to what we have with true positive false positives and so forth And so we got the sensitivity and the specificity of our study This is box plots With age and so basically by decade and so you can see A trend as well in the loss of vessels in the superficial capillary plexus and deep capillary plexus with age So after separating it into decades we then tried another classification algorithm And this time we got 35.9 percent separating it by different decades And so that's a little harder to do. We had nine different decades Options that I could do so it was still better than chance, but not as good So we were up against giants here Google had 67,001 of its studies and then the age range was 40 to 59 So a lot narrower and a lot easier to decide And then the ipax is actually just a little bit about it is actually a tele-retinal service And so they service over 300 clinics worldwide. And so it looks at all those images Their mean absolute error was small So the mean absolute error is predicted the age and then it was off by plus or minus three years And then for us we were off by plus or minus eight years, which was a lot less than standard deviation Which was about 14 So just to prove that OCT angiography will not prove but show that OCT angiography was probably the best study to use This is a saliency map and so where the algorithm that google used focused its attention and you can see The age really focused on the vessels and most ophthalmologists agreed that that green area was kind of focusing on the vessels all right, so OCT angiography um actually has been shown in other studies and it's kind of been debated in the past But it has been showing a rare refaction in the vessels or a decrease in the vessels almost annually by 0.4 percent the reduction in perfusion can Can contribute to other disease entities and so understanding that is important For us our sample size was a limiting factor I know 1800 doesn't sound like a limiting factor, but it was and then Just using machine learning as an analytical tool. I think can help drive a point So I want to end with this is a quote from shakespeare. The eyes are the window to your soul I don't think we've found any evidence for this yet But the growing body shows that perhaps it is the window to your age gender blood pressure Maybe even severe Thank you So there's no question artificial intelligence less machine learning is going to be a big part of our future Again increasingly being used and I think that A lot of physicians kind of fear that area, but there was a very interesting study Done in pathology in which When you have a whole series of large screenings That artificial intelligence did better than the average pathologist, but What was the best of all is if artificial and it was dramatically better Picked out the five most atypical areas of every specimen so definitely look at these And so what you do is is that When when when you're trying to pick a needle out of a haystack, we only have five pieces of hay That's much easier than if you have thousands of pieces hay and when you're screening a lot You sometimes go into this mode where where you're kind of numbed over you're not thinking and so By by honing in and just saying these are the places. Look it dramatically improved better than than any diagnostic criteria So I think they'll be an interesting symbiosis And a lot of mundane things will be removed and a lot more serious will be passed on But there's a lot of stuff that's going to be taken care of. I mean for most skin lesions The artificial intelligence is already way better than the average dermatologist and picking those things out But when it comes to actually treating it, I mean, that's that's where dermatology obviously is going to shine Yeah, I agree with you. I mean, I'm not an evangelist for AI I don't think it's actually going to take over but combining the two can be quite powerful That would be a very big part of us. No question. I mean, I mean already we're talking about a project now We have in which we're going to use this Program because frankly for a diabetic one of our biggest problem has they're just not getting screened In the best time to catch them when they see their primary care physician And the picture is taken and AI looks at it And and it says you need to be seen that's where the sensitivity is what's really strong Not not classifying it but making sure that all those people Get it and get in and get seen when we when it says that there there is retinopathy Because like I said the sensitivity you said was like 96 97 percent I think those are the kinds of things we'll see extremely important in our future Yeah Yeah, that's that's a great question So I guess one of the things that we did is I guess we were a little more nimble So the input variables for google was 22 million where ours was 19. So we kind of chose OCT and geography because we really thought it was related to age And so I guess being more nimble and kind of choosing your battles can help Very good. Thank you Yeah, it's not too bad. I actually picked up machine learning Um these past few years and most of the programming for this was actually done here. So Thank you