AbstractsMathematics

Learning-based Segmentation for Connectomics

by Thorben Kröger




Institution: Universität Heidelberg
Department: The Faculty of Mathematics and Computer Science
Degree: PhD
Year: 2014
Record ID: 1118041
Full text PDF: http://www.ub.uni-heidelberg.de/archiv/16834


Abstract

Recent advances in electron microscopy techniques make it possible to acquire highresolution, isotropic volume images of neural circuitry. In connectomics, neuroscientists seek to obtain the circuit diagram involving all neurons and synapses in such a volume image. Mapping neuron connectivity requires tracing each and every neural process through terabytes of image data. Due to the size and complexity of these volume images, fully automated analysis methods are desperately needed. In this thesis, I consider automated, machine learning-based neurite segmentation approaches based on a simultaneous merge decision of adjacent supervoxels. - Given a learned likelihood of merging adjacent supervoxels, Chapter 4 adapts a probabilistic graphical model which ensures that merge decisions are consistent and the surfaces of final segments are closed. This model can be posed as a multicut optimization problem and is solved with the cutting-plane method. In order to scale to large datasets, a fast search for (and good choice of) violated cycle constraints is crucial. Quantitative experiments show that the proposed closed-surface regularization significantly improves segmentation performance. - In Chapter 5, I investigate whether the edge weights of the previous model can be chosen to minimize the loss with respect to non-local segmentation quality measures (e.g. Rand Index). Suitable w are obtained from a structured learning approach. In the Structured Support Vector Machine formulation, a novel fast enumeration scheme is used to find the most violated constraint. Quantitative experiments show that structured learning can improve upon unstructured methods. Furthermore, I introduce a new approximate, hierarchical and blockwise optimization approach for large-scale multicut segmentation. Using this method, high-quality approximate solutions for large problem instances are found quickly. - Chapter 6 introduces another novel approximate scheme for multicut segmentation  – Cut, Glue&Cut  – which is based on the move-making paradigm. First, the graph is recursively partitioned into small regions (cut phase). Then, for any two adjacent regions, alternative cuts of these two regions define possible moves (glue&cut phase). The proposed algorithm finds segmentations that are { as measured by a loss function { as close to the ground-truth as the global optimum found by exact solvers, while being significantly faster than existing methods. - In order to jointly label resulting segments as well as to label the boundaries between segments, Chapter 7 proposes the Asymmetric Multi-way Cut model, a variant of Multi-way Cut. In this new model, within-class cuts are allowed for some labels, while being forbidden for other labels. Qualitative experiments show when such a formulation can be beneficial. In particular, an application to joint neurite and cell organelle labeling in EM volume images is discussed. - Custom software tools that can cope with the large data volumes common in the field of connectomics are a prerequisite for the…