Turetken E., Benmansour F., Andres B., Pfister B.H., Fua P. (2013), Reconstructing Loopy Curvilinear Structures Using Integer Programming, in Conference on Computer Vision and Pattern Recognition
, IEEE, Portland, OR.
Türetken E., Benmansour F., Fua P. (2012), Automated Reconstruction of Tree Structures using Path Classifiers and Mixed Integer Programming, in Conference on Computer Vision and Pattern Recognition
, IEEE, Providence, RI.
Türetken Engin, González Germán, Blum Christian, Fua Pascal (2011), Automated reconstruction of dendritic and axonal trees by global optimization with geometric priors., in Neuroinformatics
, 9(2-3), 279-302.
Turetken Engin, Benmansour Fethallah, Andres Bjorn, Glowacki Przemyslaw, Pfister Hanspeter, Fua Pascal, Reconstructing Curvilinear Networks using Path Classifiers and Integer Programming., in IEEE transactions on pattern analysis and machine intelligence
Full reconstruction of neuron morphology is essential for the analysis and understanding of their functioning. In its most basic form, the problem involves processing stacks of images produced by a microscope, each one showing a slice of the same piece of tissue at a different depth.So far, most commercial products provide sophisticated interfaces to reconstruct dendritic trees and rely heavily on manual operations for initialization and re-initialization of the delineation procedures. As a result, tracing dendritic trees in noisy images remains a tedious process. It can take an expert up to 10 hours for each one, which represents a significant bottleneck in neuroscience research on neuron morphology.Automated techniques have been proposed but are designed to work on very high quality images in which the dendrites can be modeled as tubular structures. They are unlikely to succeed in lower quality images where, due to the underlying neuron structure, irregularities in the dyeing process, and other sources of noise, the filaments often appear as an irregular series of blobs. Since very high resolution images take a long time to acquire and require extremely expensive equipment, such as confocal microscopes, the ability to automatically handle lower resolution and noisier ones is therefore required to make these techniques more accessible. Ideally, the painstaking and data-specific tuning that many existing methods require should also be eliminated.In preliminary work, we have proposed a simple Bayesian approach to providing this ability. It relies on tree-optimization methods for automated delineation of dendritic trees, which is already beyond what state-of-the-art techniques can produce automatically. By contrast to these methods, we do not postulate an a priori model but our algorithm learns a dendrite appearance model that is specific to each neuron we process. That gives us a handle on the differences between neurons, whose individual appearance is not always predictable.In the project we now propose, we will improve upon this approach by- training steerable filters, which are computationally very effective, to distinguish dendrite voxels from others;- introducing a tensor-voting scheme to connect them into full dendritic trees;- incorporating domain knowledge into our algorithm to disambiguate difficult situations and validate trees returned by the previous step. These will be integrated into the NeuroTracer software package currently in use at EPFL's Brain and Mind Institute, which will allow much faster processing of microscopic imagery and allow neuroscientists to collect larger databases. This will contribute to a better understanding and modeling of dendrite morphology, which is a key component of EPFL's BlueBrainProject.Furthermore, modeling elongated structures is a problem of fundamental relevance in many scientific disciplines ranging from medicine to cartography. And since, such structures appear at many different scales: Micrometers in neurons, centimeters in blood vessels, and kilometers in road-tracking problems. The learning-based techniques we will develop in this project should carry over to these other disciplines.