An efficient conditional random field approach for automatic and interactive neuron segmentation

Mustafa Gokhan Uzunbas, Chao Chen, Dimitris Metaxas
2016 Medical Image Analysis  
We present a new algorithm for automatic and interactive segmentation of neuron structures from electron microscopy (EM) images. Our method selects a collection of nodes from the watershed merging tree as the proposed segmentation. This is achieved by building a conditional random field (CRF) whose underlying graph is the merging tree. The maximum a posteriori (MAP) prediction of the CRF is the output segmentation. Our algorithm outperforms state-of-the-art methods. Both the inference and the
more » ... aining are very efficient as the graph is tree-structured. The domain of neuron segmentation requires extremely high segmentation quality. Therefore, proofreading, namely, interactively correcting mistakes of the automatic method is a necessary module in the pipeline. Based on our efficient tree-structured inference algorithm, we develop an interactive segmentation framework which only selects uncertain locations for a user to proofread. The uncertainty is measured by the marginals of the graphical model. Only giving a limited number of choices make the user interaction very efficient. Based on user corrections, our framework modifies the merging tree and thus improves the segmentation globally. This young topic named Connectomics (Sporns et al., 2005), has shown significant progress. It can 4 provide new insights into the relation between the brain structures and its function (Helmstaedter 5 and Mitra, 2012). Being able to model this relationship is believed to provide deeper understanding 6 for the principal reasons of serious brain diseases such as mental illnesses and learning disorders. 7 To obtain a complete wiring, scientists must image the brain at nanometer (nm) resolution and 8 reconstruct neuronal arbors, synapses, and glial cells. There are two notable EM imaging tech-9 niques that can provide sufficient information for reconstruction of a nervous system. i) Serial 10 section transmission electron microscopy (ssTEM) imaging can be used for reconstruction of a 11 complete nervous system of small organisms such as C. elegans or Drosophila larva and deliv-12 ers anisotropic volume with large section thickness (Briggman and Bock, 2012); ii) Focused ion 13 beam based serial section (FIBSEM) imaging is used to acquire sub volume of adult brain tissue 14 at higher resolutions with isotropic resolution in 3 dimensions (Knott et al., 2008). 15 With the developments in imaging techniques, computing power and data storage systems, to-16 day, neuroscientists can acquire large datasets in the GB-TB range. Thus, large areas of tissues can 17 be processed and meaningful neuronal circuits can be extracted by segmenting the arbors. Seg-18 mentation of arbors can be obtained by highlighting cytoplasmic membranes that separate adjacent 19 neuron cells. In fact, the whole idea of reconstruction of neuronal circuits relies on accurate detec-20 tion and segmentation process, thus small merge or split errors on membrane segmentation would 21 make the results useless and the interpretation wrong (Jain et al., 2010a; Moritz Helmstaedter, 22 2011). However, with data sets this size, manual analysis is no longer feasible.This challenges 23 the computer vision and machine learning community to develop accurate and efficient techniques 24 for neuron cell segmentation. There has been significant effort in automated methods (Jain et al., 25 2010b, 2011); but automated segmentation of EM images faces significant challenges including 26 • identical textures between adjacent neurons; 27 • inhomogeneous appearance within cells (mitochondria and vesicles); 28 • and high variation in shape (elongated, twisted) of structures. 29 Currently, manual or semi-automated interactive systems are still necessary in practice since Con-30 2 nectomics requires extremely accurate partitioning and state-of-the-art automated approaches can-31 not provide satisfactory results. In order to achieve a satisfying quality, human experts have to 32 proofread, namely, to manually correct the results of automated methods. Thus, semi-automated 33 tools which can address large data sets with reduced user interaction and low complexity are also 34 of huge interest (Sommer et al., 2011; Chklovskii et al., 2010; Straehle et al., 2012; Kaynig et al., 35 2013; Liu et al., 2014a). In this work, our goal is to provide a learning based segmentation model 36 which provides accruate automated neuronal circuit reconstruction and efficient interactive proof-37 reading mechanism. During interactions, human input is needed only at the most critical locations 38 so that the interaction burden is reduced. 39 We propose a CRF based method whose underlying graph is constructed by the output of wa-40 tershed transform (Beucher and Meyer, 1992). The watershed transform partitions a given image 41 into superpixels by simulating a water flooding of the landscape of a given scalar function, e.g. 42 the gradient magnitude or the likelihood of each pixel being the boundary. These over-segmented 43 489 tomics. Current Opinion in Neurobiology 20, 653 -666. 490
doi:10.1016/j.media.2015.06.003 pmid:26210001 fatcat:vuolpfytq5gqdpoofwn2kb2twy