Inferring Semantic Geometric Models of Biological Structure from Image Data

Dr. Kobus Barnard
School of Information: Science, Technology and Arts
Tuesday, February 19, 2013 - 4:00pm
Marley 230

I will introduce a top-down Bayesian modeling approach to extract biological
structure, with a key application being quantifying morphology. I will emphasize
how this approach is distinct from traditional image analysis methods. The
approach focusses on 3D models that represent both the semantics (e.g., identity
of parts), the topology of the assembly of parts, and the geometry of the parts
and the assembly. From here, we construct likelihood functions that
statistically explain the data, given the model. Finally, inference (fitting
models to data) is considered separate from modeling. Since we emphasize
modeling, we need to pay the price of challenging inference which we implement
with somewhat problem specif Markov chain Monte Carlo sampling strategies. I
will describe the application to inferring the structure of Alternaria (joint
work with Barry Pryor) as well as Arabidopsis (joint work with Ravi Palanivelu).