University of Cambridge > > Microsoft Research Cambridge, public talks > Frameworks for segmentation, classification, and nonstationary image processing with applications to disease assessment

Frameworks for segmentation, classification, and nonstationary image processing with applications to disease assessment

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Microsoft Research Cambridge Talks Admins.

Abstract: This talk presents several frameworks for image analysis that solve ubiquitous challenges for computer vision. The first framework addresses the dynamic segmentation of 3D deformable objects. For robust segmentation of objects undergoing complex motion, a key step is the proper implementation of conservation laws that govern shape deformation. I show how to reconstruct 3D geometry using a model from data-fusion approach and how to implement physically sound, compressibility constraints for real materials. The concepts are generally applicable to segment deformable shapes imaged over time, and I illustrate them to segment the heart in cardiac images. The second framework targets lower level image processing challenges including: denoising and motion tracking. I recast these problems in a nonstationary setting and show which properties vary over space and how to utilize that information to optimize image processing results. For denoising, I show how to construct a new, generic, nonstationary denoising framework that preserves more true structure for better image interpretation. For motion tracking, I show how the motion of objects with characteristic local line patterns can be best recovered by nonstationary weighting of Fourier domain information based on local content in the spatial domain. The third image analysis framework addresses probabilistic classification. This framework applies to problems where the primary geometric constraint is difficult to express concisely with a conservation law, but can be recovered probabilistically by learning from training data. The framework entails several steps including: data registration to form an atlas space, feature extraction, and probabilistic model construction. I will demonstrate the framework with two different approaches for the model construction step. First, using a Bayesian learning approach with a Markov Random Field, I will show how the framework enables a 10 fold increase in the number of structures that can be automatically labeled in brain MRI . The method has received FDA approval for neuroanatomical structure quantification and its applications include disease assessment of: Alzheimer`s, Huntington`s, Parkinson`s and Schizophrenia. Second, using a discriminative learning approach with the Random Forest, I show how the framework can be applied to extract biometric data about a person from an image of their face. Applications include security and surveillance tasks.

This talk is part of the Microsoft Research Cambridge, public talks series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2019, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity