%0 Generic
%A Bailoni, Alberto
%C Heidelberg
%D 2021
%F heidok:30271
%R 10.11588/heidok.00030271
%T Deep Learning for Graph-Based Image Instance Segmentation
%U https://archiv.ub.uni-heidelberg.de/volltextserver/30271/
%X Neuroscientists have been developing new electron microscopy imaging techniques and generating large volumes of data to reconstruct the complete neural wiring diagram of an organism's central nervous system. The sheer size of these volumes makes manual analysis infeasible. A fundamental step towards this goal is the automated segmentation of neural tissue images. This thesis presents new efficient deep learning methods for image instance segmentation and their applications to neuron segmentation.    Related work on instance segmentation focuses on training an accurate edge detector (represented by  a deep learning model) to predict transitions between different object instances in an image. In this thesis, we propose novel graph partitioning algorithms that can efficiently process these edge predictions and produce an instance segmentation. We specifically focus on partitioning algorithms for signed graphs with both positive and negative edge weights. By using signed graphs, the partitioning algorithm can find a previously unspecified number of instances without requiring the user to manually specify additional parameters (e.g., a tunable threshold).    In this thesis, we introduce a simple and efficient graph partitioning algorithm, the Mutex Watershed, and prove its relation to the NP-hard multicut/correlation clustering optimization problem. We then propose a generalized framework for agglomerative graph clustering algorithms, called GASP, and prove that the Mutex Watershed is one of the algorithms covered by it. This unifying framework allows us to conveniently study both theoretical and empirical properties of the algorithms it describes. When combined with the predictions of a deep neural network, some of the algorithms in the framework constitute a segmentation pipeline that achieves state-of-the-art accuracy on the CREMI neuron segmentation challenge without requiring to tune domain-specific hyper-parameters.    Finally, this thesis proposes a new bottom-up instance segmentation method for large-scale volumetric images. The approach predicts single-instance segmentation masks across the entire image, one for each pixel, in a sliding window style. All masks are decoded from a low dimensional latent representation, which results in a memory-efficient pipeline. The method achieves competitive results on the CREMI neuron segmentation challenge and is considerably robust to noise due to  prioritizing predictions with the highest consensus across overlapping masks.