CS seminar: Medical image segmentation with transformers
This event is in the past.
11:30 a.m. to 12:20 p.m.
Speaker
Chengyin Li
Graduate Research Assistant
Wayne State University
Abstract
Medical image segmentation is an important task in medical image analysis that aims to delineate structures of interest in images. Recent advances in deep learning and particularly convolutional neural networks (CNNs) have led to significant improvements in segmentation accuracy. However, CNNs usually have limitations in global context learning capability, which may be required for more challenging medical image segmentation tasks. In this talk, we explore the use of attention-based transformers for medical image segmentation. Transformers allow modeling of long-range dependencies in images via self-attention, overcoming limitations of CNNs in capturing global context. I will start by introducing how to use Transformers to perform 2D based prostate segmentation with CT images. Then, I will delve into the application of both convolution and transformer-based operations for 3D based multi-organ segmentation tasks and discuss the robustness issue with using different loss-functions. Lastly, I will introduce some ways to leverage the visual foundation models such as SAM to enhance the medical image segmentation task.
Biography
Chengyin Li is currently a fifth-year Ph.D. student in the Department of Computer Science at Wayne State University. He received his bachelor's degree from Nanjing University of Science and Technology, and a master's degree from University of Chinese Academy of Sciences. His research mainly focuses on Medical Image Applications, Trustworthy AI, and Visual Foundation Models (VFMs). He has successfully contributed to multiple AI research papers that have been accepted for publication in conference venues and journals, including MICCAI, ECML, Medical Physics, and Pediatric Research. Additionally, he has been actively engaged in research at Henry Ford Health System, with a primary emphasis on enhancing the performance and robustness of medical image segmentation tasks.