Loic Le Folgoc, Microsoft Research, Cambridge UK

Title:

Learning structure in complex data: Bayesian models, discriminative models and the models in between for medical image analysis.

Abstract:

The amount of raw medical scans available to us increases rapidly, but expert manual annotations often remain scarce and costly. I will present approaches that leverage the latent spatial structure and rich image semantics to generalize better from small annotated datasets, despite variability introduced by subject anatomies, acquisition protocol & imaging quality. We will cover applications to motion tracking and segmentation tasks. I will unabashedly range from Bayesian modelling techniques to auto-context forest architectures, before exploring forest-based message-passing models that seamlessly integrate fully automatic & user-assisted capabilities.