By Dave DeFusco
In artificial intelligence and healthcare, there鈥檚 a big problem that doesn鈥檛 get enough attention: a drought of good data. Deep learning鈥攖he kind of AI that powers everything from voice assistants to diagnostic imaging鈥攏eeds massive amounts of data to work well but in healthcare, getting access to that kind of data is easier said than done.
Hospitals and medical centers often can鈥檛鈥攐r won鈥檛鈥攕hare patient images like X-rays, MRIs or CT scans. Privacy concerns and strict regulations mean that many AI researchers simply don鈥檛 have enough high-quality, varied medical images to train their models. And without those datasets, it鈥檚 nearly impossible to build AI that can diagnose diseases across a broad spectrum of people and conditions.
A team of researchers from the Katz School's Department of Graduate Computer Science and Engineering has taken on this challenge with a powerful combination of creativity and cutting-edge computing. They鈥檝e created a new dataset and AI tool that could dramatically change how medical images are made and used in the future鈥攚ithout compromising patient privacy.
鈥淪ince hospitals wouldn鈥檛 share enough data, we decided to build our own. We collected over 250,000 medical images from open-source databases鈥攅verything from brain scans to lung X-rays to animal images鈥攁nd called the collection MedImgs,鈥 said Lakshmikar Polamreddy, a lead author of the study, Katz School Ph.D. student in mathematics and 2023 graduate of the M.S. in Artificial Intelligence.
Polamreddy was joined in the research by current students in the M.S. Artificial Intelligence: Sheng-Han Yueh, Deepshikha Mahato, Shilpa Kuppili, Jialu Li and Kalyan Roy.
This new dataset includes 61 disease types and 159 total categories of images from both humans and animals. That variety is important, because AI models learn better when they see more examples, especially of rarer conditions. With the dataset ready, the researchers developed a new image generator called the Leapfrog Latent Consistency Model (LLCM). This model is built on top of a powerful class of AI called diffusion models, which are known for creating incredibly realistic images鈥攖hink art, photos or even deepfakes.
But while traditional diffusion models are powerful, they鈥檙e also slow and computationally expensive. Creating just one high-resolution image can take dozens or even hundreds of steps. For medical applications where time, cost and efficiency matter, that鈥檚 a problem. That鈥檚 where the 鈥渓eapfrog鈥 idea comes in.
In the world of physics and math, leapfrog algorithms help solve complex equations quickly by jumping over unnecessary steps鈥攈ence the name. The team applied this method in a clever way: by solving a mathematical model of the image generation process in a more direct and efficient way, skipping the usual back-and-forth that diffusion models go through. Instead of taking 20, 50 or 100 steps to generate an image, the LLCM can create high-quality 512脳512-pixel images in as few as one to four steps.
鈥淭hat鈥檚 a huge improvement,鈥 said Roy. "It means doctors, researchers and AI developers could someday generate medical images almost instantly to test new ideas, train algorithms or even simulate rare diseases for study鈥攁ll without needing real patient data.鈥
To see how well their model worked, the researchers compared it to some of the biggest names in image generation: Stable Diffusion, Dreambooth and a previous model called Latent Consistency Model (LCM). They used a common measure of image quality called the Fr茅chet Inception Distance. Lower scores mean more realistic, higher-quality images. The LLCM blew away the competition, especially at earlier steps, showing that it could generate great images faster than the others. LLCM鈥檚 images looked much more like real medical scans, even with fewer steps.
To see if their model could handle totally new types of images, the researchers tested it on an unseen dataset of dog cardiac X-rays. It performed exceptionally well, even outshining the best AI models currently available. That鈥檚 important because it shows that LLCM isn鈥檛 just memorizing images鈥攊t鈥檚 learning to generalize, to synthesize realistic images it hasn鈥檛 seen before.
鈥淥ne of the coolest things about LLCM is that it can be fine-tuned,鈥 said Dr. Youshan Zhang, senior author of the study and assistant professor of artificial intelligence and computer science. 鈥淭hat means hospitals, labs or even veterinary clinics could customize it with their own small sets of medical images and generate thousands of realistic versions in just a few clicks. That would be a game changer for training diagnostic tools鈥攅specially in places where patient data is limited or inconsistent.鈥
