Elsevier

NeuroImage

Volume 23, Issue 4, December 2004, Pages 1283-1298
NeuroImage

New methods for the computer-assisted 3-D reconstruction of neurons from confocal image stacks

https://doi.org/10.1016/j.neuroimage.2004.06.047Get rights and content

Exact geometrical reconstructions of neuronal architecture are indispensable for the investigation of neuronal function. Neuronal shape is important for the wiring of networks, and dendritic architecture strongly affects neuronal integration and firing properties as demonstrated by modeling approaches. Confocal microscopy allows to scan neurons with submicron resolution. However, it is still a tedious task to reconstruct complex dendritic trees with fine structures just above voxel resolution. We present a framework assisting the reconstruction. User time investment is strongly reduced by automatic methods, which fit a skeleton and a surface to the data, while the user can interact and thus keeps full control to ensure a high quality reconstruction. The reconstruction process composes a successive gain of metric parameters. First, a structural description of the neuron is built, including the topology and the exact dendritic lengths and diameters. We use generalized cylinders with circular cross sections. The user provides a rough initialization by marking the branching points. The axes and radii are fitted to the data by minimizing an energy functional, which is regularized by a smoothness constraint. The investigation of proximity to other structures throughout dendritic trees requires a precise surface reconstruction. In order to achieve accuracy of 0.1 μm and below, we additionally implemented a segmentation algorithm based on geodesic active contours that allow for arbitrary cross sections and uses locally adapted thresholds. In summary, this new reconstruction tool saves time and increases quality as compared to other methods, which have previously been applied to real neurons.

Introduction

Metric analysis of neurons is indispensable to address (i) the complicated relation between physiology (function) and morphology (form) of neurons by computational approaches (Borst and Haag, 1996, De Schutter and Bower, 1994, Häusser et al., 2000, Koch et al., 1982, Segev and Rall, 1998), (ii) the characterization of cell types and the investigation of neuronal development by statistical analysis of the morphology (da Fontoura Costa and Velte, 1999, Libersat and Duch, 2002, Mizrahi et al., 2000, Uylings et al., 1986, van Pelt et al., 1989), and (iii) the relation of neuron surfaces to other structures in three-dimensional space (Belichenko and Dahlström, 1995, Gray and Weeks, 2003, Hiesinger et al., 2001, Jankowska et al., 1995, Lamotte d'Incamps et al., 1998, Wouterlood et al., 2002) (for an overview, see also da Fontoura Costa et al., 2002). Depending on the goal, a description of the neuron has to comprise the center lines and radii, the topological structure including branching and end points, the order of segments,1 or an exact reconstruction of the surface.

Generally, in order to obtain morphometric measurements, reconstructions of neurons should fulfill the following requirements: (i) sufficient accuracy must be accomplished, (ii) topological constraints based on the assumption of a treelike structure must be fulfilled, (iii) the relevant measurements must be represented explicitly allowing direct access, and last but not least, (iv) the amount of the user's time and effort should be reasonably small.

The goal of decreasing the necessary expense of user interaction often acts contrary to that of ensuring the accuracy and the topological correctness of the result. Consequently, available commercial software tools provide methods either for automatic segmentation or manual reconstruction. One of the state-of-the-art software tools of the latter category is Neurolucida (MicroBrightField, Inc., Williston, VT). Unfortunately, the manual tracing of complex dendritic trees is overly tedious. If the user tries to reduce the input actions, the result suffers from inaccuracy and abrupt changes of thickness or center line direction.

Automating the reconstruction process is difficult due to noise and the partial volume effect. Noise is generated by different sources in the confocal microscope and the partial volume effect occurs, when the volume, which is assessed for a measurement, contains labeled and unlabeled tissue. The strength of the latter depends on the point spread function (PSF). Both phenomena cause overlapping intensity histograms of the interior and exterior. In particular, fine structures have low contrast to the background, which may be decreased additionally by inhomogeneous staining of the cell (see Fig. 1).

Several software tools featuring automated reconstruction procedures are available. To the best of our knowledge, all of them work with threshold-based segmentation methods. Here we confine our discussion to the FilamentTracer (formerly called NeuronTracer) included into the program Imaris (Bitplane, Zürich). It provides three modes for the reconstruction of neurons: one works fully automatic, one semiautomatic, and the other fully manual. The semiautomatic tracing allows the user to draw along the branches in a two-dimensional projection and automatically computes the position in the third dimension. We discussed the principle problems of manual tracing with respect to Neurolucida above. The automatic reconstruction is based on a hysteresis method using two thresholds. Voxels with intensity values above the high threshold are marked as interior with high confidence (say class inHi), those between the high and the low threshold are marked as interior with low confidence (say class inLo) if there is a connection via other inLo voxels to an inHi voxel. Voxels falling below the low threshold are marked as exterior. The two thresholds are chosen by the user. Afterwards, the segmented volume is reconstructed with generalized cylinders. The result can be manually corrected.

The hysteresis method is an improvement over single threshold segmentations but still suffers from the same limitations. The appropriate thresholds depend on the staining intensity of a branch, which in turn results from its true diameter (see Fig. 1). If the user optimizes the thresholds to a specific diameter, structures with other diameters are miscalculated. In case of neurons with a limited variation of segment diameters, and therefore staining intensities, the hysteresis method improves the completeness of image segmentation. However, extensive variation in the staining intensities of neuronal structures makes the choice of appropriate hysteresis values, which include the complete topology of neuronal arborization difficult or even impossible. Therefore, this automatic reconstruction procedure is useful for fast visualization and extraction of the rough topology, whereas morphometric measurements are imprecise, in particular for neurons with a wide range of diameters. A second source of inaccuracy in the FilamentTracer is smoothing, which can be applied to the result of the reconstruction. It neglects the image data; hence, abrupt changes of the axis direction or the radii are leveled out regardless of actual data values.

Thus, the demand for the reconstruction of fine structures near the spatial resolution limits of the microscope is not answered by existing automated methods. In order to increase the efficiency of the reconstruction process without losing any quality, we developed a system that highly reduces the time investment by automatic computational methods. It assists the manual reconstruction process by (i) fitting automatically the center lines, radii, and the surface of a neural branch between points given by the user, and (ii) providing an intuitive and convenient user interface which allows immediate access to the result. The combination of the two features yields the highest possible accuracy in practice. On the one hand, the automated fitting relieves the user from most of the necessary user input, thus accelerating the process and eliminating subjective estimations and imprecision due to the user's haste or exhaustion. The reconstruction process is made mostly independent from the user's knowledge. On the other hand, the user has full control as an option allowing correction at regions where computation appears to fail.

This article is organized as follows. In the next section, we review related approaches for the detection of tubular structures in three-dimensional images. In the Methods section, we introduce the model for neural structures and describe the methods that fit the model to the data. After showing example results from real neurons in the Results section, conclusions are drawn in the Conclusions section.

Section snippets

Image analysis background

The task of reconstructing a neuron from raw three-dimensional image data can be split into two subtasks: First, a suitable representation (a model) of the reconstructed neuron has to be defined and then an efficient algorithm has to be designed, which transforms the image into a description based on this representation, that is, which fits the model to the data. According to the requirements discussed in the Introduction, the neuron should be described in a geometric way, that is, by a graph

Methods

The skeleton of the neuron is reconstructed as follows. The branching and the end points and, if necessary, some points in between are set manually by the user. This initialization roughly approximate the shape of the neuron and provides its topology. It is then fitted to the data with the automatic methods described in the Skeleton and generalized cylinders section. Initialization and fitting of segments can be iterated. Due to a low signal-to-noise ratio or an improper initialization and

Results

We applied the neuron reconstruction technique described in this paper to stacks of confocal images of cultured astrocytes, sensory neurons, inter-, and motoneurons. Fig. 7a shows a maximum intensity projection, and Figs. 7b and c the reconstruction of a dendritic tree of a motoneuron (MN5) of a Manduca sexta.7 The image in Fig. 7

Conclusions

In this paper, we present a semiautomatic method for high-quality three-dimensional reconstructions of most complex neurons from confocal image stacks, which can be used for diverse purposes. The skeleton gives the structural description of the neuron with high accuracy and almost arbitrary high sample density of center lines and radii while reducing considerably the user's effort, that is, the necessary quantity and quality of input actions. The automatic surface reconstruction achieves

Acknowledgments

This work was supported by the Federal Government of Germany (S. Schmitt, M. Scholz, K. Obermayer BMBF 0310962) and the Deutsche Forschungsgemeinschaft (J.-F. Evers SFB 515/A7; C. Duch SFB 515/A7, DU331/2-3).

References (38)

  • S.M. Pizer et al.

    Zoom-invariant figural shape: the mathematics of cores

    Comput. Vis. Image Underst.

    (1998)
  • Y. Sato et al.

    Three-dimensional multi-scale line filter for segmentation and visualization of curvilinear structures in medical images

    Med. Image Anal.

    (1998)
  • I. Segev et al.

    Excitable dendrites and spines: earlier theoretical insights elucidate recent direct observations

    Trends Neurosci.

    (1998)
  • H.B. Uylings et al.

    The metric analysis of three-dimensional dendritic tree patterns: a methological review

    J. Neurosci. Methods

    (1986)
  • K.A. Al-Kohafi et al.

    Rapid automated three-dimensional tracing of neurons from confocal image stacks

    IEEE Trans. Inf. Technol. Biomed.

    (2002)
  • S.R. Aylward et al.

    Initialization, noise, singularities, and scale in height ridge traversal for tubular object centerline extraction

    IEEE Trans. Med. Imaging

    (2002)
  • T.O. Binford

    Generalized cylinder representation

  • H. Blum

    A transformation for extracting new descriptors of shape

  • A. Borst et al.

    The intrinsic electrophysiological characteristics of fly lobula plate tangential cells: I. passive membrane properties

    J. Comput. Neurosci.

    (1996)
  • Cited by (0)

    View full text