Learning Context Cues for Synapse Segmentation in EM Volumes

Abstract

Abstract. We present a new approach for the automated segmentation of excitatory synapses in image stacks acquired by electron microscopy. We rely on a large set of image features specifically designed to take spatial context into account and train a classifier that can effectively utilize cues such as the presence of a nearby post-synaptic region. As a result, our algorithm successfully distinguishes synapses from the numerous other organelles that appear within an EM volume, including those whose local textural properties are relatively similar. This enables us to achieve very high detection rates with very few false positives.

Learning Context Cues for Synapse Segmentation

“…We describe here our approach, which was first introduced in [5]. Let x ∈ X = [0, 1] W ×H×D be an EM volume of width W , height H and depth D. Voxels are indexed by i ∈ {1, ..., W × H × D}, and the location of each voxel is designated i ∈ N 3 .…”

“…Note that, in contrast with our first approach [5] and with the work of [20], we opted for a reduced set of channels. An important observation regarding our framework lies in the fact that our features sum the response of a filter inside a box: In the case of linear filters and boxes larger than a single voxel, these sums can be directly computed in the original image channel, up to a scale factor and additional negligible filter border effects.…”

“…Several experiments confirmed this observation, showing no performance loss when using the reduced set of channels w.r.t. the full set used in [5]. Further experiments allowed us to eliminate the Hessian Eigenvalue channel, which was found to be uninformative for our framework as well.…”

Multi-Class Multi-Scale Series Contextual Model for Image Segmentation

“…Several supervised methods also have been proposed for object segmentation in EM images such as convolutional neural networks [10] and series of ANN [12] for membrane detection or [20], [30] for mitochondria segmentation or [31], [32] for synapse segmentation. However, these frameworks target only one object of interest and to our knowledge, they do not use intra-class information to give a coherent segmentation of multiple objects.…”

Refining Mitochondria Segmentation in Electron Microscopy Imagery with Active Surfaces

“…These statistics require not just object detection, but also very accurate surface structure delineation. State-of-the-art image segmentation algorithms such as [6,1,7,10] produce reasonably good localization results, in that they are able to detect most instances of the object they are searching for, and provide a rough outline. However, they often fail to accurately define the detailed boundary surface of the object in question to the precision required for accurate geometric measurements, especially surface area.…”