It's my pleasure to introduce our next speaker, Professor Ulis Bagsy. Professor Bagsy is a faculty member at the Center for Research in Computer Vision and the SAIC Chair Professor at the Computer Science Department in the University of Central Florida. His research interests are artificial intelligence, machine learning, and their applications in biomedical and clinical imaging. Dr. Bagsy. Hello, everybody. My name is Ulis Bagsy from the Center for Research in Computer Vision, University of Central Florida, Orlando. I'm going to talk about algorithms for medical AI in general, and the talk is called an alphabet soup. I'm very happy to be here for the second time last year. It was also a fantastic summit. First, I'm going to talk about very briefly the evolution of the algorithms, AI algorithms, and then the current weaknesses and the drawbacks of the AI algorithms. I will try to mention and show examples for why we need an explainable AI for creating a trust between the user and the physician and the machine learning algorithms. And then as a case example, how we can increase the trust and robustness by putting the physician in the loop when we develop deep learning algorithms. And then I will show some case examples and conclude my talk. So as you already know, neural networks now date back to several years ago. In 2014, the first, you know, the idea was there, algorithm was there, but because of the lack of, you know, the sources, power sources like GPU and optimization algorithms, more than two, three layers of deep learning algorithms were not possible. In 2014, the winner of the ImageNet algorithm is AlexNet, as you can see here. It was with, you know, the eight layers of deep learning architectures. Since then, we start seeing more advanced algorithms are coming with, you know, the deeper and deeper architectures. At that time, this was a deep architecture. And then VGG11 came, then VGG13, which means like 13 layers, 11 layers, then 16 layers, 19 layers, and it start increasing as time goes on. More complex and more advanced neural networks architecture comes into play for solving more complex problems in almost every field. In 2015, you see the architecture who, which was winning the ImageNet competition was having more than 150 layers. And here you see the, once the networks were becoming deeper and deeper, algorithms are becoming more and more advanced and advanced and the accuracy of those systems are getting higher. At the same time, the operations are, you know, you see the 155 million operations happening. And so there is another field comes into, subfield comes into field, like how to reduce computations and make the systems more eco-friendly. So the, you know, the while, you know, these evolutions of these algorithms are still taking place and we, every day, we notice that actually the new algorithms are coming to solve an important problem in different fields, in medicine, in autonomous driving cars, in other fields. There are, there are many also drawbacks in the current AI systems. AI fails badly. Most of the algorithms are non-robust. For example, here you see a robot passport checker, which is rejecting Asian man's application because eyes are considered close. Another one is because of the gender, race, and all these, you know, the biases existing in the algorithm or in the data, you know, there are algorithmic biases, there are data biases, a robot judge, a beauty contest, and don't select women with dark skin. So this is one of the problem. So from the medical imaging example here, you see a dermal toscopic image of a benign cyst where the classification algorithm is showing the, this is a benign. When you put a little bit of noise, you see the overall picture doesn't really change this photograph. However, the machine learning system decides oppositely. It shows malignant. This was actually published in science in 2019. And there are, for example, here, you see the, when you rotate the image and put them into the, you know, the machine learning system, it gives you a wrong diagnosis again. So especially, you know, these, these fails are very risky in high risk applications, right? So it's not really desirable. So, you know, what comes into mind is, again, is an old notion, however, so when you have a system, machine learning system, which classifies, let's say, this cat, it should come with explanations so that you can actually trust. And this is going to increase also the robustness of the algorithm, right? Because you will know this is cat because it has fewer risk, gloves, and all other things. But the current explanation of the machine learning system actually doesn't have this. The other thing, actually, when we describe is the explainable, actually, this is, this is something really a domain specific notion. So there cannot be all purpose definition for explainability. For a cat, you know, the classification, it is, there is different explanations. For medical AI, there are different explanation modules. And for autonomous cars, the explainability components are all different. So interpretable or explainable to whom, right? So in our case, the end point is usually the physician, right, the physician should be able to understand the decisions so that the algorithms can be trusted. There are actually explainable AI approaches so far, but most of these approaches are actually is a post hoc methods. So usually actually showing where, you know, the pixels for the imaging applications, I give example, the pixel which contribute maximally to the prediction are considered like where the algorithms are learning. So it's like a saliency map, it shows actually, you know, we are learning here or we are learning here for the decision. This another way is the occlusion visualization where you just get rid of some of parts, some part of the images and see the probability plot for the classification. When there is a big jump, you understand actually those parts are important parts that, you know, the fewer filters learn in deep learning. However, many of these approaches are post hoc. They are not inherently explainable. So there are artifacts, signal loss and all other common problems in visual maps. You see in the right hand side, the correct label are just actually here seen, but your attention map or your, you know, the saliency map show that the filters learn from outside lung region for the lung disease classification. So still those methods are not perfect and they are not inherently explainable. We really need inherently explainable algorithm, which gives reason. And then we need to make sure that the algorithms are robust and also not fragile with simple attacks. And that is true for like all high risk application, especially, right? Non-application, fine. All high risk application and including GI applications, we really need to have this. I'm going to show one example to increase the robustness and trust in current machine learning algorithms, which is actually, is again, old, old phenomena. So we're going to include the human in the group. And in our case, the human is our expert physicians. So I think it's very simple. Two brain is better than one brain. So human and AI combined is better than AI. You know, it's going to complement the drawbacks of those, you know, the AI system. So one way actually we are going to do that actually, there may be many reasons, many different way of doing that. In order to provide the one-to-one realistic interaction between machine learning algorithms and the physician is to use an interface. For interface, actually, we are using eye tracking technologies. In the past, we were using these glasses. Now we are using desktop monitor, desktop eye trackers, where radiology or physician is not really bothered by anything, just doing their own routine stuff while doing that actually we track their eyes, where they look at it and how they manipulate the images and analyze the images, quantify them. So we have the visualization system, MIPAP, it can be any PAC system in radiology or in other fields like including GI. These videos of colonoscopy, for example, can be in the same way it can be used. And we have the eye tracker system. And then it goes to the C-CAT, which is collaborative AI algorithm. So true collaboration is happening. And it can be in dark environment, you know, the bright environment. So it's not really bothering the physician. It is real time and very realistic. So it can be even multiple, you know, the monitors, it can be done. Again, this is the old style glasses. We don't have that now. We have a monitor, desktop monitor. You see multiple images can be even, you know, analyzed here as well. So this is the prostate, for example, cancer screening experiments we have done at NIH. So from the algorithm behind the settings, actually, it's pretty straightforward setting. So we collected gaze data real time. The data is really huge. So we use unsupervised clustering techniques to reduce the data without losing a lot of information. And then we use this reduced data in the deep learning settings. So these are coming from gaze patterns, which means it is true attention. It's the true attention regions. Unlike the, you know, the famous attention mechanism is being mentioned by several computer scientists. We are using true attention here because it directly comes from the gaze patterns. Then actually your task can be anything, right? So you can do segmentation. You can do detection. You can do classification like, you know, the diagnosis. So you see here and then one example for a lung cancer screening study. I want to show in the left hand side the CT. These heat maps are showing where the radiologists look at it. And once the screening is done, you see this green and red region. And red region on the left is something like the radiologist didn't look at it. When this is happening, the deep learning algorithm is always running in the background. So once it is finished, deep learning algorithm finds if there is any nodules and show this region to radiologists saying that actually you cover this region. Everything's fine, but that region is not covered. So there is, you know, we found that. So there is this true interaction and it's really mutually benefit to each other. And we can even understand actually how radiologists or physicians are moving from one image to another image and what information can be combined. We can understand from gaze patterns as well. And we use this in a quantitative way, not really qualitative evaluation, but we put them into the deep learning system for the interaction. This is for another, like this is so useful that actually we use in many other places. Here we are segmenting brain tumors just by looking at them. And eye tracker device you see here and in the background of the workstation, there is always an algorithm like UNET or something else ongoing. You just look at the high level and then we segment, for example. So since data collection is very, very important and labeling and it's very costly, I think this is very innovative way of collecting data without using additional time, but just doing the routine evaluation of those scans or, you know, during the surgery for GI application, this can be really done in real time. And you see different segmentation, different imaging modalities. Once rough segmentation, ground labeling is done, we can improve them and use it, especially in medical imaging. This is a big problem, like we don't have really millions of images to label, but this is one way to actually go towards that direction. So just before actually finishing my talk, I, you know, I really like this, the roadmap published in ideology 2019, it shows the, you know, foundational research on AI for medicine, medical imaging. So, first of all, of course, we need to improve data quality, right? So data may have, you know, the biases and algorithm biases need to be removed as well. This is another. So the data, we need to increase the data quantity, but also quality and then the labeling. So the labeling, you know, is very, very expensive. Therefore, our innovative technology for eye tracking based, you know, the data labeling can be used. And then, so we need to, you know, deal with the clinical reality, how we are going to actually, you know, translate algorithms into the clinics. So we need to increase the trust. And this is very important. The robust algorithm should come and wherever it's necessary, we need to really use, you know, the physician's input, right? And the software use, understand how AI works. This is interpretability and explainability. We need to understand model interpretation, but we need to understand also the reasoning, why those diagnoses are made or why those prognostic decisions were made, right? Or surgery, you know, the modeling and predictions. We need to always have, you know, the ties to the reasons. And so for algorithm training, we need more robust algorithm, not fragile. And then that can be done with, you know, incorporating more high-level information from the physician in the loop. And also like new level of algorithms are coming like capsule algorithms instead of CNN that we are also using. And we are following this evolution of those algorithms. And from the understanding point, reasoning point of view, we need to have more explainable and more intelligent algorithm, less artificial algorithms. In conclusion, you know, we, so we need to increase the transparency. We need to build a trust for medical AI, and we need to make sure that, you know, we have fair models. And interpretation of models will help us, you know, understand how the learning is working. And this is the way, you know, at least one way we can facilitate incorporation of such, you know, adaptation of such technologies in the clinics. Thank you very much for listening. I'm happy to get any, you know, question in the discussion section. Thank you.

AI algorithms

medical applications

explainable AI

biases in algorithms

errors in medical diagnosis