Abstract

New technologies for acquiring massive amounts of medical image data give rise to an ever-increasing demand for effective approaches for medical image analysis tasks. In recent years, deep learning (DL) approaches have yielded remarkably high-quality solutions for numerous medical imaging applications, largely outperforming traditional image analysis methods. Comparing to natural scene images, medical image analysis faces several different challenges. Commonly, DL methods rely on a large amount of labeled data for model training. While natural scene images are usually 2D, medical images can be 2D, 3D, and even higher dimensional. In particular, 3D medical images are widely used in basic research and clinical practice. However, 3D medical image analysis presents big challenges to DL methods. First, 3D medical images can be of very large sizes (e.g., billions of voxels), and thus incur high computational costs. But, current GPUs are of limited memory for implementing 3D DL models. Furthermore, few efficient automatic techniques for labeling 3D images are available. Since in general, only trained medical experts can label medical images effectively, medical image annotation is a highly costly and labor-intensive process (even for 2D images). Therefore, how to attain sufficient good quality labeled image data for DL model training while significantly reducing annotation efforts of medical experts is a big bottleneck to the successful development and deployment of DL methods for medical imaging applications.

We present new DL-based approaches for medical image analysis tasks (segmentation, classification, denoising, etc). We show that it is often not enough to simply apply DL methods alone to tackle medical image analysis problems. Thus, our approaches are based on combinations of DL methods and algorithmic techniques (e.g., topological data analysis). For example, our sparse annotation schemes judiciously select the most representative or valuable samples to label. Actually, the problem of finding an optimal subset of samples (as sparse labeled data) to cover or represent an entire image dataset is an NP-hard problem, which can be solved approximately with guaranteed good quality. Our approaches achieve high performances with efficient costs. We present experimental results on various datasets to demonstrate the applicability of our approaches on medical image analysis problems.

Bio.

Dr. Danny Z. Chen received the B.S. degrees in Computer Science and in Mathematics from the University of San Francisco, California in 1985, and the M.S. and Ph.D. degrees in Computer Science from Purdue University, West Lafayette in 1988 and 1992, respectively. He is a Professor in the Department of Computer Science and Engineering at the University of Notre Dame. Dr. Chen’s main research interests include computational biomedicine, biomedical imaging, machine learning, data mining, computational geometry, algorithms, and VLSI. He has worked extensively with biomedical researchers and practitioners, published many papers in these areas, and holds 8 US patents for technology development in biomedical applications. He received the US NSF CAREER Award in 1996 and the 2017 PNAS Cozzarelli Prize of the US National Academy of Sciences. He is a Fellow of IEEE and AAAS, and is a Distinguished Scientist of ACM.

All faculty and students are welcome to join.