时间:3:00pm – 5:00pm, April 28, 2016
地点:310 room, Office building (行政楼310会议室)
题目: Alternating Direction Methods in Medical Image Segmentation
Abstract:Recently, optimization of high-order and fully connected graphical models hasdrawn tremendous research interests. Such hard-to-optimize functions arisenaturally in a breadth of vision, learning and medical imaging problems. Forinstance, they can be very powerful in the context of image segmentation,boosting substantially the performances of state-of-the-art supervised learningmethods such as convolutional neural networks. This is particularly the casewhen training annotations are limited/weak or when image data undergo very poorcontrasts. In this talk, I will discuss some recent developments in thisdirection, focusing on a general and powerful alternating direction (ADMM)framework for optimization. I will show the potential of this framework inmedical image segmentation and discuss its key technical aspects, using variousillustrations and applications. In particular, I will show how ADMM can (1)deal effectively with difficult high-order fractional terms; and (2) distributesystematically the computations for a general class of pairwise functions.
Bio:Dr Desrosiers is professor at the ETS Quebec Univeristy, and co-director of theLIVIA laboratory. Prior to joining ETS, he was postdoctoral researcher underthe supervision of George Karypis at the University of Minnesota, USA. Hereceived his Ph.D. in Applied Mathematics from Ecole Polytechnique in Montreal.His recent research focuses on applying machine learning and mathematicalmodeling to various image analysis and computer vision problems.