Item

Abstract

Recent methodological advances (SBFSEM, see [1]) have made it possible for the first time to scan large
volumes of nervous tissue rapidly and automatically at electron microscopic level. If individual neurons
could be segmented in these image stacks, neuronal circuits with all synapses could be reconstructed. Due
to the huge amount of image data it is not practical to do the segmentation by hand, and methods for fast
and automated segmentation are needed.
Contrary to previous work [2], we approach the segmentation problem as one of finding local edge probabilities
in the 3-dimensional image block. First, 3-dimensional Fourier transforms are computed on local
blocks of 4x4x4 voxels. Next, an unsupervised clustering algorithm (K-Means) is used to group the
Fourier blocks into typical image features. This reduces the computational cost, as further computations
can be performed on the smaller set of image features rather than on the original data. It also helps to
reduce image noise. Then, we fit generated ground-truth data, which contain edges of known position and
orientation, to the image features. For each voxel we can then compute an edge probability from the
blocks overlapping it. A flood-filling algorithm can then rapidly segment the image into its components.
This method could be easily extended to incorporate further constraints, such as edge continuity. The
method is quite fast, since much of the computation can be done once for a certain type of images and
then stored and applied to new images; for example the clustering into features and association of edge
probability patterns to these features. This probabilistic framework offers great potential for automated,
efficient and robust segmentation of neurons in electron-microscopic image stacks.