세미나안내

세미나안내

Domain Generalization for Computer Vision and Medical Imaging Analysis

2021-04-23
  • 1,621
황성재 교수 (University of Pittsburgh) / 2021.04.07

[Biography]

Dr. Seong Jae Hwang received his B.S. in Computer Science from the University of Illinois at Urbana-Champaign in 2011, M.S.E. in Robotics from the University of Pennsylvania in 2013, and Ph.D. in Computer Sciences from the University of Wisconsin-Madison in 2019. He is currently an assistant professor of Computer Science and Intelligent Systems Program in the School of Computing and Information at the University of Pittsburgh. His research is focused on developing statistical machine learning and deep neural network methods for analyzing imaging modalities in computer vision, machine learning, and medical imaging. On the technical side, he develops algorithms for cross-sectional and sequential data from small to large scales with statistical machine learning and deep learning models. On the application side, his interests range from neuroscientific discoveries including understanding the pathological progression of Alzheimer’s disease to machine learning/computer vision applications.

[Abstract]

Modern neuroimaging studies often combine data acquired from multiple scanners and experimental conditions. Such collection of data may contain substantial technical variability associated with scanner effects where images from different scanners possess non-biological, systematic variability. Analogous to the batch effect correction in other domains, my colleagues and I investigate the problem of neuroimaging data harmonization to enable accurate analysis of data from multiple sites or scanners. Interestingly, from a machine learning perspective, such multi-site/scanner neuroimaging data also poses an empirical challenge: a model trained on one scanner performs poorly when tested on another scanner. I will discuss our recent works by first characterizing this problem with a pre-existing notion of domain generalization from computer vision. Then, I will show how this technique can significantly improve the generalizability of a brain lesion detection model on new, unseen neuroimaging data.

ZOOM : https://zoom.us/j/8978217407?pwd=d1pOZmF1OWlseEdZRVBpV3VuSkl3dz09

ID : 897 821 7407

PW : 1nTQDY

목록