Raphael Sznitman (École Polytechnique Fédérale de Lausanne - EPFL,Switzerland)
With their ability to image with isotropic resolution of up to 4nm per pixel, Scanning Electron Microscopes (SEM) have become invaluable tools for studying intra-cellular structures and model organelles such as mitochondria, synapses, and vesicles. To acquire image stacks from 3D tissue samples, SEMs scan a rectangular region of the block face several times and average the results to form a clear image of the tissue surface. A thin layer from this surface is then milled and the process repeated for the extent of the tissue sample. Unfortunately, doing so for the smallest of samples can take days.
In this talk, we will consider the problem of finding visual targets quickly to improve SEM throughput for experimental design setups. As such, the problem is closely related to Active Learning and Optimal Control. Within a Bayesian framework, we consider the task of making a limited number of sequential observations, each one evaluating an imperfect classifier of chosen cost and accuracy on a chosen region of the tissue surface, in order to minimize an objective that combines the entropy of the posterior distribution and the cost of observations. We will show that a Bayes-optimal, one-step lookahead solution for any arbitrary time horizon can be computed and implemented. We will then show how this result can be used in the context of SEMs to allow speed-ups in time for image acquisition.
Raphael Sznitman received his B.Sc. in cognitive systems from the University of British Columbia in 2007. Following this, he went on to study at the Johns Hopkins University, where he received his M.Sc and PhD in computer science in 2011. He is currently, a postdoctoral fellow at the École Polytechnique Fédérale de Lausanne (EPFL,Switzerland) where he works in the computer vision laboratory. His research interests lie in the fields of computer vision and machine learning with applications to biomedicine and health.