Automated three-dimensional detection and counting of neuron somata

Abstract

We present a novel approach for automated detection of neuron somata. A three-step processing pipeline is described on the example of confocal image stacks of NeuN-stained neurons from rat somato-sensory cortex. It results in a set of position landmarks, representing the midpoints of all neuron somata.

In the first step, foreground and background pixels are identified, resulting in a binary image. It is based on local thresholding and compensates for imaging and staining artifacts. Once this pre-processing guarantees a standard image quality, clusters of touching neurons are separated in the second step, using a marker-based watershed approach. A model-based algorithm completes the pipeline. It assumes a dominant neuron population with Gaussian distributed volumes within one microscopic field of view. Remaining larger objects are hence split or treated as a second neuron type.

A variation of the processing pipeline is presented, showing that our method can also be used for co-localization of neurons in multi-channel images. As an example, we process 2-channel stacks of NeuN-stained somata, labeling all neurons, counterstained with GAD67, labeling GABAergic interneurons, using an adapted pre-processing step for the second channel.

The automatically generated landmark sets are compared to manually placed counterparts. A comparison yields that the deviation in landmark position is negligible and that the difference between the numbers of manually and automatically counted neurons is less than 4%. In consequence, this novel approach for neuron counting is a reliable and objective alternative to manual detection.

Introduction

The number of neurons in the brain and their varying density between different brain regions is thought to be a fundamental determinant of brain function (Donaldson, 1895, Williams and Herrup, 1988, Williams and Rakic, 1988). During the last 60 years, great effort has been made to estimate neuron densities quantitatively, first addressed by Abercrombie's article “Estimation of Nuclear Population from Microtome Sections” (Abercrombie, 1946). The estimation of absolute numbers of neurons, densities or rates of density change in neuron populations is usually based on random, sparse sampling methods (Cragg, 1967, Rockel et al., 1980) such as stereology (Sterio, 1984). These methods determine cell densities by inspecting a representative sub-volume of tissue and extrapolating the obtained density values to a reference volume. Usually these density values are given with an accuracy of about 10% (Beaulieu, 1993) for large anatomical units, such as primary visual (V1), somato-sensory (S1) or motor cortex (M1). However, the deviation between densities in previous studies is much larger. For instance in V1 a variety of densities values are reported (40,000 (Cragg, 1967, Knox, 1982, Werner et al., 1982), 52,000 (Beaulieu, 1993), 75,000–80,000 (Peters et al., 1985, Warren and Bedi, 1984) neurons per mm³). It is hence difficult to determine density changes within or between neuron populations, or across functional sub-units, such as a cortical column in S1 (Helmstaedter et al., 2007). In consequence, it would be favorable to count the absolute number of neurons in large volumes (∼0.5 mm³ for a cortical column in S1 (Meyer et al., in preparation-b)) and hence derive the detailed three dimensional neuron distribution of the brain area of interest.

Recently available three-dimensional imaging techniques (mosaic/optical-sectioning confocal laser scanning or mosaic/optical-sectioning widefield microscopy) and suitable neuronal stains opened new possibilities for the determination of neuronal densities within entire volumes. Neuronal stains, like NeuN (Gittins and Harrison, 2004, Kumar and Buckmaster, 2007, Mullen et al., 1992, Wolf et al., 1996) labeling all neuron somata, or GAD67 (Kaufman et al., 1991, Meyer et al., in preparation-b, Muzio et al., 2002, Staiger et al., 2002) labeling GABAergic interneuron somata, as well as genetically encoded labels of specific neuron populations in transgenic mice or drosophila (Akemann et al., 2004, Luo et al., 2008) allow in principle the quantitative determination of density differences between neuron populations at high level of detail (e.g. between or within cortical layers).

Several neuron counting and detection methods have been reported, both manual and automated ones. The obvious disadvantage of manual neuron detection, apart from possible subjectivity, is the amount of time needed for neuron counting. In consequence, automated accurate detection and segmentation of neurons from microscopic images has been extensively studied (Liu et al., 2008). In general, these algorithms can be divided into three categories: threshold-based (Wu et al., 2000, Wu et al., 1995), watershed-based (Lin et al., 2003, Lin et al., 2005, Malpica et al., 1997, Nilsson and Heyden, 2005, Vincent and Soille, 1991) and model-based approaches (Chang and Parvin, 2006, Li et al., 2006, Lin et al., 2007, Lin et al., 2005, Raman et al., 2007, Ranzato et al., 2007).

None of the three described automated algorithm categories yield satisfying results for the here presented sample data. This is due to the fact that model-based methods have comparatively better specificity in detecting the targets, i.e. such methods find all objects satisfying the model-shape but only those. In contrast, threshold- or watershed-based approaches display relatively better detection sensitivity, i.e. they find all objects, but usually result in incorrect numbers (Liu et al., 2008), e.g. for our sample data touching, densely clustered neurons are counted as one, resulting in about 20% less neurons.

Here we present an automated 3D neuron counting approach that combines all three approaches to a novel high-throughput system for detection of neuron somata (alternatively called neurons throughout this article). The system is described on the example of confocal image stacks of NeuN-labeled neurons from rat primary somato-sensory cortex (S1). The slightly adapted method for co-localization in multi-channel images is described on the example of NeuN neurons counterstained with GAD67. This allows to measure the fraction of GABAergic interneurons among all neurons (Meyer et al., in preparation-b). We will show that our algorithms are robust and adjustable to different microscope magnifications (40×, 63×), and various imaging (confocal, two-photon, widefield) and staining (NeuN, GAD67, Ca²⁺-sensitive dye) techniques.

The presented processing pipeline consists of three steps and is summarized in Fig. 1. The goal of the first threshold-based step is to create a binary image separating foreground (i.e. stained neurons) from background. It consists of a number of image processing steps, including compensation for imaging or staining artifacts such as bleaching, shading or uneven uptake of the stain, and binarization by local thresholding. This step is data-specific. We will show that the pre-processing algorithm for confocal stacks of NeuN-stained neurons can be systematically adapted to different staining or imaging methods. This is illustrated by two examples: widefield stacks of NeuN-stained neurons and in vivo two-photon (2p) stacks of neurons containing Ca²⁺-sensitive dye. This threshold-based approach is usually not sufficient to detect the true number and position of neurons. High neuron densities and limited microscope resolution result in clusters of neurons that cannot be separated by the local threshold step. The first processing step is therefore regarded as a pre-processing step that guarantees a similar input to the second (watershed-based) and third (model-based) processing steps. The implementation of the latter two steps is independent of the data type.

In the second watershed-based step, clusters of neurons which are connected by narrow links are separated by a morphological filtering process, resulting in an image of distinct watershed regions (3D objects of connected foreground voxels, identified by a label number), and ideally representing individual neurons. Some clustered neurons appear however like a single, large and uniformly stained neuron. The morphological filters are not capable of splitting such clusters into distinct watershed objects.

The third, model-based processing step addresses this problem. We assume a single dominant neuron population within the image stacks with a Gaussian-distributed neuron (i.e. soma) volume. The mean neuron volume and its variance are calculated from a volume histogram of the watershed regions. Undivided clusters are then split according to their volume, assuming that it has to be an integer multiple of the mean soma volume. An additional advantage of this constraint is that its parameters are not specified by the user but automatically calculated during the image processing. In a correction step, eventually present spatially separated neuron sub-populations with larger mean volumes are investigated and remain unsplit.

We also present an extension of our method to multi-channel image stacks for co-localization of counterstained neurons. Here NeuN-labeled neurons were counterstained with GAD67 in order to measure the fraction of GABAergic interneurons among all neurons (Meyer et al., in preparation-b).

In the results section, we compare our detection algorithm with manual counts. We show that the automated system reproduces manual counts with more than 90% precision. The 10% deviations originate from detection of ambiguous objects, meaning touching neurons that could for instance be counted as one or two neurons. However, manually counted but automatically missed neurons (false negative, FN) as well as automatically detected but manually missed neurons (false positive, FP) comprise about 5%, respectively. This compensatory effect results in average relative counting differences of approximately 1%. The average absolute deviation in neuron numbers was determined as less than 4%. In addition the deviations in landmark positions proved to be negligible. Further, the automated approach is much faster, reducing manual labor of approximately 4 h per stack to a few minutes and additional computing time of about 1 h.