Abstract text (incl. figure legends and references)

Introduction

Traditionally, neuron reconstructions from serial EM micrographs and synapses localizations are done manually, which is extremely time consuming. Currently used, semi-automatic selection tools for segmenting are faster compared to manual selection. In order for those tools to work properly, prior to segmentation, noise reduction is necessary, while keeping membrane integrity and structural features of synapses preserved at the same time.

Objectives

Here we introduce a solution, which includes pre-processing of electron micrographs. Pre-processed images are suitable for the use of more advanced, fast selecting tools, which significantly shorten the amount of time needed for segmentation.

Materials and methods

Stacks of serial electron micrographs of identified neurons of the locust, Locusta migratoria were produced by serial block face scanning electron microscopy. This was followed by multiple image processing steps on these stacks by Image J software.

First preprocessing step: A histogram matching algorithm was used to homogenize the initial brightness levels within the stack, followed by a low pass Fourier transform band-filter to isolate small structures like cell-membranes. Further a rolling ball algorithm was used to give the cell membranes a higher contrast and thereby the stack was set up for an automated Otsu threshold. The thresholded stacks were analyzed with the particle analyzer implemented in ImageJ to filter out small irregular components present in the original stack not belonging to the cell membranes. In parallel, the particle analyzer was used once more with settings that enabled keeping the large particles such as mitochondria. The result was subtracted from the result of the first round with the particle analyzer.

Second preprocessing step: mitochondria, representing an obstacle during and after segmentation, where the target. For this purpose the histogram - matched stacks were filtered with a median filter after a contrasting step. Finally a Huang´s auto threshold was used to get a binary image (that contained the mitochondria) which was subtracted from the first preprocessing step.

Results

The first step successfully removed many structures inside the cells, but mitochondria attached to the membranes were still visible, compromising the results of the following segmenting steps. This can slow down fast selection tools subsequently used for segmentation (such as "Blow" in Amira). (Fig. 1B)

During the second preprocessing step most of the structures inside the cells were deleted. However, at certain locations the membranes appeared discontinuous. Nevertheless combining both steps with fast selection tools significantly enhanced segmentation efficiency compared to a traditional manual approach. (Fig. 1C)

Conclusion

Pre-processing can improve the accuracy and efficiency of segmenting algorithms.