Dartmouth Events

Abstract:

There exists an abundance of neural network techniques in computer vision that achieve impressive performance on various visual processing tasks. Most of these methods require large and fully-supervised training datasets, impeding their applicability in the medical imaging domain. In this talk, I discuss the limitations of large-scale supervised training for medical image analysis, and propose several approaches to address these challenges. Specifically, I show how synthetic data can improve computational model performance when real datasets are small or not available. I also discuss how weakly-supervised and unsupervised learning can circumvent the reliance on large-scale supervision. The proposed new methodology is evaluated on various computer vision and medical imaging benchmarks and medical imaging data.

Bio:

Dr. Elena Sizikova is a Moore Sloan Faculty Fellow in the Center for Data Science, New York University (NYU). She received her BA in Mathematics and Computer Science from the University of Oxford, UK in 2013. She completed her PhD in Computer Science at Princeton University in July 2019, where she was NSF Graduate Research Fellow in the 3D Vision Lab advised by Professor Thomas Funkhouser. During her PhD studies, she spent time as a research intern at Siemens Healthineers and Adobe Research. She has received best paper awards at the ECCV 2016 Workshop on Virtual and Augmented Reality (VARVAI) and the EUROGRAPHICS 2016 Workshop on Graphics and Cultural Heritage (GHC). In recognition for her work, she was selected as a 2020 Rising Star in Engineering in Health in the School of Engineering and College of Physicians and Surgeons at Columbia University. Elena's research focuses on developing new computational methods and algorithms in computer vision which aim to address pressing challenges in medical imaging, biomedical research, and more generally visual understanding tasks (see https://esizikova.github.io to learn more about her work).