A software package for
Automated Neuronal Morphology
3DMA-neuron is part of the 3DMA software package for analyzing biphase two- and three-dimensional images. It has the capability to determine automatically detailed analysis of dendritic length and characteristics of dendritic spines [7], as well as determining automatically branching morphology of neurons [8, 10], requiring only a limited number of parameters. Further, these two functionalities have been integrated, so that both the branching morphology as well as the fine scale structure of a neuron can be analyzed when imaged at sufficiently high resolution. The high degree of automation and relative speed in computational time make 3DMA-neuron well suited to obtaining detailed morphology of multiple neurons, allowing comparison among different samples.
The program runs on the Linux operating system, and requires either Motif or the LessTif libraries.
The image analysis algorithms (for both the dendritic spine detection
and the dendritic branching detection) consist of three general steps:
image segmentation, extraction of the medial axis of the neuron phase,
and analysis of the medial axis and neuron phase. These steps are
shown below, with appropriate illustration of both these applications.
Also described are the morphological measurements obtained after the appropriate
dendritic structures are identified.
Morphology of Dendritic Spines
The principal components of the fine structure analysis consist of spine detection, morphological characterization of each spine, and dendritic radius determination. The algorithms used for spine detection/characterization have been described in detail elsewhere [7] and have four main components: segmentation; dendritic backbone extraction; spine detection and merging; and measurement of morphological characterizations.
Segmentation
The greyscale intensity of each voxel is an integer from 0-255 (for 8-bit images).In the GUI, there are two ways to segment the image: using simple segmentation and using indicator kriging. For simple segmentation (Fig. 4), the user must enter a single threshold: voxels having intensity above this threshold are assigned to the neuron phase; remaining voxels are assigned to the background phase. Indicator kriging [12] is a locally adaptive segmentation method, which requires that a subpopulation of voxels of each phase be positively identified a priori. This can be done by manually establishing a window of intensity values delimited by two thresholds. Intensity values above the upper threshold are assumed neuron, those below the lower threshold are assumed background tissue. The classification of remaining voxels is estimated by a two-point correlation function incorporating spatial information from the a priori classifications. See Fig. 1.
Fig. 1: xy-projections
of (a) a CLSM image of part of the dendritic tree from a pyramidal neuron
and (b) the neuron phase (shown in black) identified after segmentation.
Construction and Trimming of the Medial Axis
The medial axis of a digitized object is the minimally connected skeleton by which the original topology and geometry of the object is preserved. The medial axis of the neuron is composed of voxels from the neuron phase which are equidistant from voxels on the neuron surface. Erosion-based algorithms are used to construct the medial axis; they are based on the work of Lee et al. [9]. The medial axis is very sensitive to surface noise; as a result, the medial axis may contain a number of small spurious paths that do not contribute to the overall geometry of the object. Such spurious paths are trimmed using a minimum length criterion. Fig. 2 shows the medial axis extracted from the raw data shown in Fig. 1(a), before (a) and after (b) spurious paths are trimmed.
Fig. 2: Shown are
xy-projections
of (a) the medial axis extracted from the neuron phase identified in Fig.1b;
(b) the medial axis, after trimming short spurs and resolving loops.
The colors represent, in rainbow scale, the diameter of the dendrite at
each medial axis voxel.
For more information on the medial axis algorithms, see the Online
Manual to 3DMA.
Dendritic Backbone Extraction
After the medial axis has been constructed and trimmed, the dendritic backbones are extracted from the medial axis by trimming short `spurs' and resolving loops. The user may select whether a dendritic backbone will be continued through a branch point. If the option to terminate backbones at branch points is not selected, a backbone is continued through a branch point according to a minimum deviation angle criterion as described in [7].
Spine Detection
Spine detection proceeds in three steps: detection
of spine components entirely detached from dendrites containing backbones,
detection of spine components attached to these same dendrites, and merging
of two or more spine components. Any dendritic phase component containing
no dendritic backbone whose center of mass is sufficiently close to a dendritic
backbone is identified as a detached spine head. Attached spines
are detected as local protrusions of a dendritic surface. To identify
spines which comprise more than one attached or detached spine component,
a final merging algorithm is performed. Fig. 3 shows the result of
spine detection on the raw data shown in Fig. 1.
Fig. 3: Shown are
xy-projections
of dendritic backbones and spines identified from Fig. 1 after the spine
detection algorithms are completed. Barely visible in this figure
is the backbone of each identified dendritic segment. The color scheme
is arbitrary.
Manual Spine Editing
As a final level of user control, the GUI provides a facility by which
the user can scan projected images showing identified spines, and by mouse
click remove spines (or dendritic segments) deemed to be false positive
signals.
Dendritic Branching Morphology
For a typical neuron image, there are three main components to the algorithms
we employ for determination of the dendritic branching morphology: segmentation
of the raw data to distinguish the neuron phase from the background phase;
identification of the dendritic structure using skeletonization; and labeling
of branch points
(and hence branch segments) with respect to branch order. These
are described in detail elsewhere [8].
Segmentation
See above for a description of available segmentation algorithms.
Fig. 4: xy-, xz-, and yz-projections of (a) raw and (b) segmented data of a pyramidal neuron. Simple segmentation has been used, in contrast to the kriging results shown above in Fig. 1.
Obtaining the Neural Skeleton
The largest connected component of the neuron phase is retained as comprising the cell; disconnected components are discarded as debris. Once this is done, both the medial axis extraction and trimming are performed as described above in the spine detection application. Fig. 5 shows the trimmed medial axis result.
Fig. 5: xy-projection
of the extracted skeleton of the neuron pictured in Fig. 4.
Labeling Branch Points
The limits of the soma are delineated manually using the graphical user interface (GUI) [8]. The skeleton inside the soma region is omitted from further analysis. We refer to each dendritic trunk leaving the soma as the root of a dendritic tree. Using the GUI, manual point and click is used to select one (or more) trees that will be identified as the apical arbor. All other trees comprise the basal arbor. Any axon collaterals which are detected can be removed manually. Dendritic branches in each tree are labeled according to a centrifugal nomenclature [14]. Looping structures in the skeleton are resolved prior to labeling. The final result is shown below.
Fig. 6: (a) Apical
and (b) basal arbors detected, after manual selection of the soma region
and apical brush. The skeleton segments designated as interior to
the soma region are shown in pink. Remaining colors, in rainbow scale,
correspond to the branch orders of the neuron branch segments.
Multiscale Morphology
We have integrated the dendritic branching and spine detection algorithms to obtain multiscale morphological analysis of pyramidal neurons. An entire pyramidal neuron from the macaque monkey superior temporal cortex has been imaged in three dimensions at high resolution (0.098 x 0.098 x 0.081 um) via confocal laser scanning microscopy, in several stacks. The various stacks were integrated into a single volume as described in [13], resulting in a 25 GB data set. Such a data set is too large to be analyzed as a single volume by the computational power accessible to most labs.
Fig. 7: xy-projection of an entire pyramidal neuron from the macaque monkey superior temporal cortex, imaged at 0.098 x 0.098 x 0.081 um resolution. Individual tiles manually selected for spine detection analysis are also indicated.
For dendritic spine morphology analysis, the data set was first subdivided into non-overlapping tiles; algorithms to detect spine characteristics were then applied to each tile (see above). To analyze the dendritic branching morphology, the global coordinates of all detected dendrite fragments are mapped into a smaller cube. This provides a representation of the dendritic structure amenable to automated branching morphology analysis (see above). An alignment procedure was applied to associate the spine data from each fine scale dendritic fragment to the corresponding neuron branch identified in the compressed structure. The result is a detailed morphometric description of the entire neuron. For more detail on these algorithms, see [16].
(a)
(b)
(c)
Figure 8: Reconstructed
(a) apical and (b) basal arbors of the neuron. In (a), the apical
shaft (shown in black), the terminal tuft (shown in red), and oblique branch
dendrites (shown in remaining colors) are identified. (c) shows a
close-up of some of the branches. The dots represent spines associated
with the dendrites.
Morphological Measurements
Recorded for each spine are: length, volume, head and neck diameters, position on the dendritic backbone, and shape classification (thin, mushroom, stubby) [11, 5]. Spines which appear disconnected are assigned a neck diameter of 0 and are designated as mushroom shaped. Spine volume is determined to be sum of the volume of all voxels contained in the spine.
Recorded for each dendritic branch are: total length, running diameter, number of spines, spine density, mean spine length, and relative populations of spines of each classification type. Dendritic diameter determination arises directly from the skeletonization procedure used to determine the backbone of each dendritic branch segment lying in each tile. The radius is measured as a function of position along any dendritic branch segment.
The branching morphology is output in SWC format [2], a standard text
file which lists basic information about the soma and each dendritic branch.
More detailed output, including summary details about the spine shapes
on each branch, can be output in NeuroML format [4, 3], a markup language
which serves as an interchange format between various software tools.
These formats allow the data to be easily incorporated into compartmental
modeling software such as NEURON [6] and GENESIS [1].
References:
[1] J. M. Bower and D. Beeman. The Book of GENESIS: Exploring realistic neural models with the General Neural Simulation System. TELOS/Springer-Verlag, 1994.
[2] R. C. Cannon, D. A. Turner, G. K. Pyapali, and H. V. Wheal. An on-line archive of reconstructed hippocampal neurons. J. Neurosci. Methods, 84: 49-54, 1998.
[3] N. H. Goddard, D. Beeman, R. Cannon, H. Cornelis, M. O. Gewaltig, G. Hood, F. Howell, P. Rogister, E. De Schutter, K. Shankar, and M. Hucka. NeuroML for plug and play neuronal modeling. Neurocomputation 44: 1077-1081, 2002.
[4] N. H. Goddard, M. Hucka, F. Howell, H. Cornelis, K. Shankar, and D. Beeman. Towards NeuroML: model description methods for collaborative modeling in neuroscience. Philos. Trans. R. Soc. B, 356: 1209-1228, 2001.
[5] K. M. Harris, F. E. Jensen, and B. Tsao. Three-dimensional structure of dendritic spines and synapses in rat hippocampus (CA1) at postnatal day 15 and adult ages: implications for the maturation of synaptic physiology and long-term potentiation. J. Neurosci. 12: 2685-2705, 1992.
[6] M. L. Hines. The Neuron simulation program. In J. Skrzypek, editor, Neural Network Simulation Environments, pages 147-163. Kluwer, Norwell, MA, 1994.
[7] I. Y. Y. Koh, W. B. Lindquist, K. Zito, E. A. Nimchinsky, and K. Svoboda. An image analysis algorithm for dendritic spines. Neural Comput., 14: 1283-1310, 2002.
[8] Y. Y. Koh. Automated recognition algorithms for neural studies. PhD thesis, Stony Brook University, May 2001.
[9] T. C. Lee, R. L. Kashyap, and C. N. Chu. Building skeleton models via {3-D} medial surface/axis thinning algorithms. CVGIP: Graph. Models Image Process. 56: 462-478, 1994.
[10] M. Maravall, G. M. G. Shepherd, Y. Y. Koh, W. B. Lindquist, and K. Svoboda. Experience-dependent changes in dendritic morphology of layer 2/3 pyramidal neurons during a critical period for developmental plasticity in rat barrel cortex. In preparation, 2003.
[11] E. A. Nimchinsky, B. L. Sabatini, and K. Svoboda. Structure and function of dendritic spines. Ann. Rev. Physiol. 64: 313-352, 2002.
[12] W. Oh and W. B. Lindquist, Image thresholding by indicator kriging. IEEE Trans. Pattern Anal. Mach. Intell. 21: 590-602, 1999.
[13] A. Rodriguez, D. Ehlenberger, K. Kelliher, M. Einstein, S. C. Henderson, J. H. Morrison, P. R. Hof, and S. L. Wearne. Automated reconstruction of three-dimensional neuronal morphology from laser scanning microscopy images. Methods 30: 94-105, 2003.
[14] H. B. M. Uylings, A. Ruiz-Marcos, and J. van Pelt. The metric analysis of three-dimensional dendritic tree patterns: a methodological review. J. Neurosci. Methods 18: 127-151, 1986.
[15] C. M. Weaver. Automated morphometry for neural cells. PhD thesis, Stony Brook University, August 2003.
[16] C.
M. Weaver, W. B. Lindquist, S. L. Wearne, and P. R. Hof. Automated
algorithms for multiscale morphometry of neural cells. In preparation,
2003.