Biography of the speaker
Dr. Dong Yang received his Ph.D. from Rutgers University in 2019 under the supervision of Professor Dimitris Metaxas, and has continued his research on medical image processing since then. He also interned at Siemens Healthineers (formerly Siemens Corporate Research) where he worked on several medical image segmentation projects. Currently he is an applied research scientist at NVIDIA. His research interests include medical image processing and clinical statistics. He has published in many major conferences and journals such as MICCAI, CVPR, ICCV, Medical Image Analysis, IEEE Transactions on Medical Imaging, etc.
Abstract
Nowadays, with technological advancements in algorithm design (such as deep learning) and hardware platforms (such as GPUs), medical image analysis has become a critical step in disease understanding, clinical diagnosis, and treatment planning. Among various tasks, image segmentation has been one of the most important medical image analysis tasks. Recently, deep convolutional neural networks have been widely applied in medical image segmentation with state-of-the-art performance. Meanwhile, Automated Machine Learning (AutoML) has also been explored in deep neural networks, aiming to further enhance model efficiency and effectiveness. However, the existing AutoML algorithms have taken on a singular perspective and focused on separate components of deep learning solutions (e.g., neural architecture, hyperparameters), which could lead to suboptimal results.
In this talk, I will take a systematic perspective, and introduce a novel method which automatically considers and estimates most, if not all, of the components of a deep neural network-based solution for 3D medical image segmentation. The proposed method can predict the relationship between different training configurations and neural networks, which can be used for comparison of solutions. Specifically, I will introduce a new search space for neural architectures and a predictor-based AutoML algorithm to accommodate the large search space. Experiments show that the proposed method can achieve state-of-the-art performance on large-scale lesion segmentation datasets compared to other existing methods in the literature. Furthermore, the proposed method has been shown to transfer efficiently to different datasets.
In the rest of the talk, I will briefly discuss other topics in medical image segmentation, such as transformer-based networks, segmentation in federated learning, segmentation in semi-supervised learning, shape priors in segmentation, etc.