Segmentation and Tracking in Bioimage Analysis,
and the Discrete Optimization Problems they engender
Abstract: In connectomics,we wish to trace each and every neurite in massive 3D images; and in developmental biology, we hope to track the position and fate of each and every cell. Both problems remain difficult, in spite of ever-improving data quality.
In this talk, I will present both our modeling efforts and progress in the associated combinatorial optimization problems. In connectomics, I will show that the NP-hard but principled multicut / correlation clustering problem affords a great boost in accuracy over deep neural networks alone. I will also introduce heuristic solvers that reach good minima in a fraction of the time required by full integer linear program solvers.
On the tracking side, I will advocate a model for *joint* segmentation and tracking. If cells can divide, this model also is NP-hard. Again, a specialized heuristic solver can come to the rescue in large instances.
The connectomics pipeline currently tops the leaderboard of the blind ISBI 2012 and the CREMI challenges, and the tracking work performed best on at least one dataset of the last Cell Tracking Challenge.
Bio: Fred develops and applies machine learning methods to solve challenging problems in bioimage analysis. He is particularly interested in active or weakly supervised learning. His group puts great emphasis on the development of user-friendly software (such as the http://ilastik.org framework) that can be used by experimentalists.
Fred studied at ETH Zurich. He is a Professor at the University of Heidelberg, a co-founder of the Heidelberg Collaboratory for Image Processing (HCI), a Visitor to the HHMI Janelia Farm Research Campus, andstill thinks of science as the greatest job on earth.