Detection of neuron membranes in electron microscopy images using a serial neural network architecture

Abstract

Study of nervous systems via the connectome, the map of connectivities of all neurons in that system, is a challenging problem in neuroscience. Towards this goal, neurobiologists are acquiring large electron microscopy datasets. However, the shear volume of these datasets renders manual analysis infeasible. Hence, automated image analysis methods are required for reconstructing the connectome from these very large image collections. Segmentation of neurons in these images, an essential step of the reconstruction pipeline, is challenging because of noise, anisotropic shapes and brightness, and the presence of confounding structures. The method described in this paper uses a series of artificial neural networks (ANNs) in a framework combined with a feature vector that is composed of image intensities sampled over a stencil neighborhood. Several ANNs are applied in series allowing each ANN to use the classification context provided by the previous network to improve detection accuracy. We develop the method of serial ANNs and show that the learned context does improve detection over traditional ANNs. We also demonstrate advantages over previous membrane detection methods. The results are a significant step towards an automated system for the reconstruction of the connectome.

Learning Context Cues for Synapse Segmentation

“…Semi-automated methods based on active contours and level sets [4], [22], [7], [31], [17], [24] as well as graphcuts [29] have achieved some measure of success on EM images. However, these methods require careful manual initialization of each object to be segmented, which is done by supplying seed points and tuning various parameters.…”

“…For instance, [36] uses vesicle detection cues to suppress false alarms on vesicle clusters that can interfere with mitochondria segmentation, while [17], [35] propose to sample features in a 2D stencil neighborhood around the pixel of interest. By allowing the classifier to measure features computed at various locations in addition to the pixel of interest, [17], [35] are able to identify membranes at regions of minor discontinuities. However, by relying on a pre-determined set of locations from which features can be sampled, these approaches strongly restrict the use of context.…”

Learning Context Cues for Synapse Segmentation in EM Volumes

“…Human experts confirm the presence of a synapse by looking nearby for post-synaptic densities and vesicles. This protocol cannot be emulated by measuring filter responses at the target voxel [4], pooling features into a global histogram [6,7] or relying on hand-determined locations for feature extraction [8,9]. To emulate the human ability to identify synapses, we design features, termed context cues, that can be extracted in any cube contained within a large volume centered on the voxel to be classified, as depicted by Fig.…”

“…Several fully automated approaches to reliable segmentation of organelles, such as mitochondria [6,7] or neuronal membranes [8,9], from 3D EM stacks have recently been proposed. However none of these methods exploit context in a meaningful way.…”

“…However none of these methods exploit context in a meaningful way. Though features are extracted in a neighborhood around the voxel of interest, they are either pooled into global histograms [6,7] or computed at predetermined spatial locations [8,9]. The resulting classifier is therefore unable to hone in on arbitrary localized context cues.…”