By Dave DeFusco
Artificial intelligence is becoming an important helper in medicine. One area where it shows great promise is medical imaging, especially helping doctors find cancer earlier and plan more precise treatment. A study by researchers in the Katz College鈥檚 Department of Graduate Computer Science and Engineering accepted for publication in IEEE Transactions on Multimedia introduces a method called Diff-MedSeg that takes a meaningful step forward in this effort.
IEEE Transactions on Multimedia is a leading journal in its field, ranked among the very best, with a high impact score of 9.7 in 2025, meaning its research is widely read and cited by scientists working in multimedia and signal processing.
The study, 鈥淒iff-MedSeg: Diffusion Model based Medical Image Segmentation with Multi-Channel Attention,鈥 focuses on a task called medical image segmentation, which teaches a computer to look at a medical image, such as a colonoscopy image or an MRI scan, and clearly outline areas of interest, like a polyp or a prostate tumor. These outlines help doctors see where potential cancer is, how big it is and how it might change over time.
This kind of work is especially important for colorectal and prostate cancer. Colorectal cancer is the third most commonly diagnosed cancer worldwide, and prostate cancer is the second most common among men. Together, they account for more than 3 million new cases each year. When doctors can find and define tumors early, survival rates can be dramatically higher鈥攗p to 90 percent in some cases.
Today, much of this outlining work is still done by hand. Radiologists may spend 15 to 30 minutes carefully tracing tumor boundaries for a single patient. Even then, results can vary widely between experts, especially when images are unclear or boundaries are hard to see. This is where AI can help.
鈥淢edical images are often noisy and inconsistent,鈥 said Dengyi Liu, lead author of the paper and a Ph.D. student in computer science. 鈥淟esions can have fuzzy edges, low contrast or unusual shapes. Our goal was to design a system that can focus on what really matters in an image while ignoring distractions.鈥
The team鈥檚 approach builds on a newer type of AI known as a diffusion model. To understand diffusion models, it helps to imagine noise being added to an image, like static on a TV screen, and then slowly removed step by step. During this cleanup process, the model learns how to rebuild meaningful structure. In Diff-MedSeg, this process is guided so that the final result is a clean, accurate outline of a tumor or organ.
What makes Diff-MedSeg different is the addition of something called multi-channel attention. Medical images are analyzed through many layers, or channels, each capturing different kinds of information鈥攅dges, textures, shapes, and more. Multi-channel attention helps the system decide which of these channels are most important at each step.
鈥淒ifferent channels carry different medical clues,鈥 said Honggang Wang, chair of the department. 鈥淏y teaching the model to pay more attention to the channels that highlight cancer-related features, we can improve accuracy, especially in difficult cases.鈥
To test their method, the researchers evaluated Diff-MedSeg on six well-known medical imaging datasets. Three involved polyp segmentation from colonoscopy images, and three focused on prostate segmentation from MRI scans. These datasets are commonly used by researchers around the world, making them a strong benchmark for comparison.
Across nearly all measurements, Diff-MedSeg outperformed existing methods, including popular tools like U-Net, newer transformer-based models and earlier diffusion-based approaches. The improvements were especially clear when images had unclear boundaries or low contrast, which often cause problems in clinical settings.
鈥淲e also ran additional experiments to confirm that the multi-channel attention component played a key role,鈥 said Julia Fang, a professor of computer science and engineering. 鈥淲hen this feature was removed, performance dropped, showing that it was central to the system鈥檚 success.鈥
While the results are promising, the researchers note that challenges remain. Diffusion models can be slower than simpler AI systems, which may limit real-time use during procedures like live colonoscopy. Future work will focus on speeding up the process, improving interpretability and testing the system across hospitals and imaging conditions. Even so, the researchers believe Diff-MedSeg points to an exciting direction for medical AI.
鈥淭his work shows that diffusion models with attention mechanisms are a powerful new tool for medical image analysis,鈥 said Ming Ma, the corresponding author of the study and an assistant professor of computer science and engineering. 鈥淲ith further development, they could help doctors detect cancer earlier and plan treatments with greater confidence.鈥
