Illinois Medicine

called Pretraining with GANs," he says. GANs (generative adversarial networks) are deep-learning models that can create and modify realistic- looking synthetic images from real images, also known as deepfakes. "GANs learned from unlabeled data how to create realistic-looking images. By training a GAN on a large amount of unlabeled data in the hospital system, we can then fine-tune small components of it later on in a much smaller labeled data set and still get very good results on the core task that we were originally interested in." Building Up and Out The third focus of Yi's lab is to build out the college's computing infrastructure. Achieving deep learning, according to Yi, requires a different type of computer structure from what is usually available. He's looking at how data is shared across labs and colleges throughout the University of Illinois Chicago. "Right now, the system was not built for high-throughput data science research, so it's a difficult task to start a project that needs a million images: You'd have to download each one by hand. Eventually, all of the medical institutions across the nation will have an improved data engineering system. But, for now, our lab is working to provide our physicians with the AI/ML tools that can best help them do their work in the most equitable way possible to benefit all patients." Illinois Medicine | 27 optimizing algorithms taken from natural-image computer vision for medical imaging. Natural images are those taken by individuals with a smartphone camera, for example. Most computer vision focuses on these types of images. "For the most part, medical computer vision currently uses the same algorithms developed by other nonmedical AI groups and hopes that it can learn on medical data just as well as on images that you might find on any smartphone," says Yi. However, Yi's core hypothesis is that algorithms developed around the use case of medical imaging will outperform ones that were developed for more general uses. "I modeled this part of the lab to improve our performance on medical data by designing our networks (or architecture or models) around medical problems," Yi explains. "Medical problems are often overlooked when developing model architectures but deserve much more respect. It's not the same thing as dog versus cat; it's much more subtle." Another core issue with medical data is the cost to annotate the data. For natural images, crowdsourcing is often used to label commonly understood classes such as automobiles and bikers. For medical data, however, a clinician's expert understanding is needed, which ends up becoming an expensive task that further burdens the healthcare system. "One way that our lab is tackling the issue of not enough labeled data is by developing a procedure T The Role of Deepfakes in the Diagnosis of Eye Disease T E C H N I C A L L E A D : Darvin Yi, PhD I N D I V I D U A L C O N T R I B U T O R : Luca Cellamare, MS '22 THE ALGORITHMS USED TO CREATE "DEEPFAKES"—or AI-generated imitations of real images—have created a world of trouble for celebrities and cybersecurity experts in recent years. Yet deepfakes can also be used to transform healthcare by training digital systems to recognize abnormalities in medical images that may indicate disease. Yi's research team is using generative adversarial networks (GANs) to build a database of ophthalmological images. According to Yi, a database of synthetic images on its own has limited utility. "But, through the process of learning how to create these synthetic images, these neural networks are able to learn more about the physiology of the retina," he says. "We can then harness that learning to help diagnose eye disease." Deepfakes for good: Yi's team built a database of fake fundus images to train AI systems to spot eye diseases.

• Cover