1: \documentclass[aps,pre,a4,twocolumn,groupedaddress,showpacs,floatfix]{revtex4}
2: \usepackage{amsmath}
3: \usepackage{amssymb}
4: \usepackage{graphicx}
5: \usepackage{epic}
6: \usepackage{eepic}
7: \usepackage{subfigure}
8: \bibliographystyle{apsrev}
9: \begin{document}
10:
11: \title{Morphological relationship between axon and dendritic
12: arborizations as revealed by Minkowski functionals}
13: \author{Marconi Soares Barbosa$^1$ and Luciano da Fontoura Costa$^1$}
14: \affiliation{$^1$Institute of Physics at S\~ao Carlos,
15: University of S\~ ao Paulo, S\~{a}o Carlos,
16: SP, PO Box 369, 13560-970,
17: Phone +55 16 3373 9858,
18: FAX +55 162 71 3616,
19: Brazil,
20: marconi@if.sc.usp.br}
21: \date{\today}
22: \begin{abstract}
23: The spatial structure of the axonal and dendritic arborizations is
24: closely related to the functionality of specific neurons or neuronal
25: subsystems. The present work describes how multiscale Minkowski
26: functionals can be used in order to characterize and compare the
27: spatial organization of these two types of arborizations. The
28: discrimination potential of the method is illustrated with respect
29: to three classes of cortical neurons.
30: \end{abstract}
31:
32: \pacs{87.17.Nn, 87.57Nk, 87.80.Pa, 89.75.-k}
33: %Morphology of nerve cells, Image analysis, Morphometry and stereology, complex systems
34:
35:
36: \maketitle
37:
38: Human intelligence is to a great extent determined by the
39: connectivity between the myriad of neuronal cells in the nervous
40: system (Cajal:1989). Biologically, these connections are
41: implemented by the \emph{synapses} between neurons, which are
42: strongly related to the shape of axons and dendrites
43: (e.g.~\cite{Costa_can:2005, pelt:2002b, Ascoli:2004}). For
44: instance, neural cells exhibiting more intricate dendritic or axonal
45: arborizations tend to promote more connections. Therefore, the
46: systematic study of the morphological properties of the axonal and
47: dendritic trees can provide essential information about the
48: connectivity of the nervous system. Because neurite outgrowth can
49: be understood as a complex dynamical system involving pattern
50: formation, biochemical regulation and electrical activity, they are
51: particularly suitable for characterization and modeling by using
52: concepts and methods from physics, computation and biology.
53:
54: As the spatial distribution of synapses is to a great extent
55: influenced by the interaction between the shape of dendrites and
56: axons, and also because they share several developmental aspects, one
57: question of particular relevance concerns the study to which extent
58: these two types of structures are related (i.e. similar or distinct).
59: Interestingly, both axons and dendrites start their lives as
60: \emph{neurites}, being differentiated during development. Along the
61: early stages of neuronal growth, which involve many outgrowths and
62: retractions, one of these neurites dominates and become the cell axon,
63: assuming distinct biomolecular composition and function. In
64: particular, axons seek actively for targets while being influenced by
65: many trophic factors including differential adhesiveness,
66: galvanotropism and chemotropism (e.g.~\cite{Sanes:2005, Cline:2001,
67: Kiddie:2005}). Consequently, the shape of axons is directly affected
68: by external field influences. At the same time, both dendrites and
69: axons share intrinsic genetic and biochemical basis, which constrains
70: neuronal growth and may imply intrinsic similarities between the
71: dendritic and axonal arborizations of a same cell or category of
72: cells. The current article addresses this important problem from the
73: objective perspective of neuronal shape quantification by additive
74: functionals~\cite{Raedt:2001}, which have been previously applied with
75: encouraging success to the characterization of dendritic morphology
76: (e.g.~\cite{Barbosa:2003a,Barbosa:2003b}).
77:
78: Integral geometry algorithms have been successfully used to
79: characterize morphologically complex patterns and structures whose
80: process of formation is not precisely known and is subject of
81: modeling~\cite{Raedt:2001}. The central procedure is the calculation
82: of intrinsic volumes or Minkowski functionals, a generalization of
83: the usual determination of volume.
84:
85: The Minkowski functionals of a body $K$ in the plane are
86: proportional to familiar geometric quantities, namely its area
87: $A(K)$, perimeter $U(K)$ and the connectivity or Euler number
88: $\chi(K)$. The usual definition of the connectivity from algebraic
89: topology in two dimensions is the difference between the number of
90: connected $n_c $ components and the number of holes $n_h$,
91: $\chi(K)=n_c-n_h$. In the Euclidean space, there is an additional
92: geometric quantity, the mean curvature or breadth. Moreover there
93: are two kinds of holes to consider: A pure hole, which is a
94: completely closed region of white voxels surrounded by black pixels
95: and tunnels. The Euler characteristic for the Euclidean space is
96: then given by
97: \begin{equation}
98: \chi(K)= n_c -n_t +n_h,
99: \end{equation}
100: where $n_t$ is the number of tunnels and $n_h$ is the number of pure
101: holes. One instance where these functionals appear naturally is
102: while attempting to describe the change in volume as the body $K$,
103: now assumed to be convex, undergoes a dilation through a parallel
104: set process using a ball $B_r$ of radius r
105: \begin{equation}
106: \label{eq:dila} V(K \oplus B_r)=V(K)+S(K)r+2\pi B(K) r^2
107: +\frac{4\pi}{3}r^3.
108: \end{equation} Generalizing to higher dimensions, the change in hyper volume is given
109: by the Steiner formula
110: \begin{equation}
111: \label{eq_dilageral}
112: v^d(K\oplus B_r)=\sum_{\nu=0}^d \binom{d}{\nu}
113: W_{\nu}^{(d)}(K)r^{\nu},
114: \end{equation}
115: where the coefficients $W_{\nu}^{(d)}$ are referred to as Minkowski
116: functionals. For instance in the Euclidean space ($d=3$),
117: \begin{align}
118: \label{eq:familiar}
119: &W_0^{(3)}(K)=V(K),\quad W_1^{(3)}(K)=\frac{S(K)}{3},\\ \nonumber
120: &W_2^{(3)}(K)=\frac{2\pi}{3}B(K),\quad W_3^{(3)}(K)=\frac{4\pi}{3}\chi(K).
121: \end{align}
122: Despite the wealth of results and continuum formulae for obtaining
123: these functionals, it is useful to resource to the discrete nature of
124: the binary images we wish to analyze by looking at the distribution
125: of voxels in a Euclidean spatial lattice. By exploring the
126: additivity of the Minkowski functionals their estimation resumes
127: to counting the multiplicity of basic building blocks that
128: disjointedly compose the object. The fundamental information needed
129: here is a relationship for the functionals of an open interior of a
130: $n$-dimensional body $K$ which is embedded into a $d$-dimensional
131: space
132: \begin{equation}\label{eq_open}
133: W_{\nu}^{(d)}(\breve{K})=(-1)^{d+n+\nu}
134: W_{\nu}^{(d)}(K), \nu=0,\ldots, d.
135: \end{equation}
136: With the absence of overlap between these building blocks and using
137: the property of additivity of these functionals, we may write for a
138: pattern $\mathcal{P}$ composed of disjoint convex interior pieces
139: $\breve{N}_m$,
140: \begin{equation}\label{eq_whole}
141: W_{\nu}^{(d)}(\mathcal{P})=\sum_{m}
142: W_{\nu}^{(d)}(\breve{N}_m)n_m(\mathcal{P}), \; \nu=0,\ldots,d\quad,
143: \end{equation}
144: where $n_m(\mathcal{P})$ stands for the number of building elements
145: of each type $m$ occurring in the pattern $\mathcal{P}$. For the
146: three-dimensional space we display in Table~\ref{table:build3d} the
147: value of Minkowski functionals for the building elements in a
148: orthogonal lattice of voxels. By using the information (with $a=1$)
149: presented in Table~\ref{table:build3d} and Equations
150: \eqref{eq:familiar}, \eqref{eq_open} and \eqref{eq_whole} we have
151: \begin{align}
152: &V=n_3, \quad S=-6n_3+2n_2, \\ \nonumber
153: &2B=3n_3-2n_2+n_1,\quad \chi=-n_3 +n_2 -n_1+n_0,
154: \end{align}
155: Where $n_3$ is the number of interior cubes, $n_2$ is the number of
156: open faces, $n_1$ is the number of sides and $n_0$ is the number of
157: vertices. So the procedure to calculate Minkowski functionals of a
158: pattern $\mathcal{P}$ has been reduced to the proper counting of the
159: number of elementary bodies of each type that compose a voxel
160: (cubes, faces, edges and vertexes) involved in the make up of
161: $\mathcal{P}$.
162: \begin{table}[htb]
163: \begin{center}
164: \begin{tabular}{llllll}\hline\hline
165: $m$& $\breve{N}_m$ &{$W_0^{(3)}$}&{$W_1^{(3)}$}&{$W_2^{(3)}$}&{$W_3^{(3)}$} \\ \hline
166: 0 & $\breve{V}$ & 0 & 0 & 0 & 1 \\ \hline
167: 1 & $\breve{L}$ & 0 & 0 & 1/2 & 1 \\ \hline
168: 2 & $\breve{F}$ & 0 & 2 & 1 & 1 \\ \hline
169: 3 & $\breve{C}$ & 1 & 6 & 3/2 & 1 \\ \hline
170: \hline
171: \end{tabular}
172: \caption{Minkowski functionals $\mathcal{M}_{\nu}^{(3)}$ for open elementary
173: open sets $\breve{N}_m$ which compose a voxel $K$(of side
174: $a=1$): $\breve{V}$ ({\it vertex}), $\breve{L}$ ({\it side}), $\breve{F}$
175: ({\it face}) e $\breve{C}$ ({\it cube}). \label{table:build3d}}
176: \end{center}
177: \end{table}
178:
179: The recent effort aimed at replicating a working micro column of
180: cortex tissue~\cite{Markram:2006} has motivated the dissemination of
181: comprehensive data bases containing different anatomical classes of
182: neuronal cells. We have considered several neurolucida data from
183: anatomically different types of cells, namely the anatomical classes
184: MC1, BTC and LBC that populate the cortex and are engaged in
185: distinct functions in the working of a neocortical micro column. The
186: morphological data was obtained from the Brain Mind Institute, EPFL
187: database~\cite{database:2003}. Exact dilations of both the axonal
188: and dendritic trees were performed until the majority of holes and
189: tunnels disappeared, i.e., when the Euler number approaches 1. This
190: implied a total number of 1000 dilations. All four Minkowski
191: functionals were calculated along such dilations, so as to obtain a
192: morphological signature for the dendritic and axonal arborizations
193: of each cell. We then conducted a morphological analysis using a
194: subset those values at selected scales, namely after 1,250, 500, 750
195: and 1000 exact dilation steps. In order to reduce the effect of the
196: size of the cells on the measurements, the functionals were
197: normalized respectively to the adequate power of the diameter of the
198: cell, i.e the volume functional was normalized by $L^3$, the surface
199: area was normalized by $L^2$ and the mean-curvature was normalized
200: by $L$. The Euler-number is a dimensionless topological measure and
201: consequently did not undergo normalization. All measures were
202: subsequently standardized, which is accomplished by subtracting the
203: average and dividing by the standard deviation of each respective
204: type of measurement (i.e. one of the 4 functionals). The
205: standardized values therefore have zero means and unit variance, and
206: most of their values are comprised between -2 and 2. Such a
207: standardization reduces the influence of the overall relative
208: magnitude of the different measurements.
209:
210: The functionals were calculated separately for the dendritic and
211: axonal trees of each neuron. For each case, the obtained
212: measurements were organized into morphological vectors in the
213: 20-dimensional space defined by the five spatial scales of each of
214: the four functionals. Therefore, given a neuronal cell represented
215: by their respective measurement vector, the morphological difference
216: between its axon and dendritic arborizations can be quantified in
217: terms of the magnitude $\mu$ of the vector corresponding to the
218: difference between the respective feature vectors (i.e. axon and
219: dendrites). Another important descriptor of the morphological space
220: which has been considered in this work is the angle $\phi$ between
221: the above difference vector and the vector which represents the mean
222: of all such difference vectors. Such a measurement provides and
223: indication of how much the axon-dendrite in each cell departs from
224: the overall prototypical case (i.e. the average).
225:
226: \begin{figure*}[htb]
227: \begin{centering}
228: \includegraphics*[scale=.3,angle=0]{distrib_labeled.png} \caption{The distribution of axonal-dendritic similarity among cells
229: in the three neuronal classes.\label{fig:distrib}}
230: \end{centering}
231: \end{figure*}
232:
233: The result of the above procedure is presented in the scatterplot
234: (i.e. the angle \emph{versus} the magnitude of difference vectors
235: for each neuron) in the composited Figure~\ref{fig:distrib}, which
236: also shows the relative densities of the magnitudes of the
237: difference vectors obtained by using a simple Gaussian fit. Recall
238: that total morphologic similarity between the axonal and dendritic
239: trees for each cell would result in a magnitude of difference vector
240: equal to zero. The scatterplot in Figure~\ref{fig:distrib} shows
241: that most cells resulted with magnitude between 0 and 2 and angle
242: between 0 and 6, with an outlier with angle near to 10 in the case
243: of the MC2 cell category. As shown by the respective densities, the
244: three classes of neuronal cells resulted with distinct dispersions,
245: with the MC2 class being characterized by the smallest variance
246: while the BTC class shows the largest variance. In order to obtain
247: a more objective indication of the separation between the three
248: neuronal cell categories, we performed the ANOVA variance test
249: (e.g.~\cite{Hogg:1987}) and obtained the results ($p-$values) shown
250: in Table~\ref{table:anova}. These values indicate that the MC2 and
251: LBC classes are those less likely to have come from the same
252: population of cells, with a respective $p-$value of only 0.0088.
253: Substantially higher values were obtained for the other pairs of
254: categories, suggesting less separability in those cases.
255:
256: \begin{table}[htb]
257: \begin{center}
258: \begin{tabular}{llll}\hline\hline
259: Class & BTC & LBC & MC2 \\ \hline
260: BTC & & 0.2676 & 0.1629 \\ \hline
261: LBC & & & 0.0088 \\ \hline
262: MC2 & & & \\ \hline
263: \hline
264: \end{tabular}
265: \caption{The resulting $p-$values obtained through ANOVA for the 3 classes of neurons. \label{table:anova}}
266: \end{center}
267: \end{table}
268:
269: Figure~\ref{fig:distrib} also shows an example of neuronal cell for
270: each of the anatomical classes, lying close to the center of the
271: respective Gaussian distribution fit. Despite the difficulty of
272: comparing the original 3D axonal and dendritic arborizations from such
273: 2D projections, the larger difference between the two arborizations
274: obtained in the case of the LBC category is still clear, as well as
275: the greater similarity observed for the MC2 cell example.
276:
277: The obtained results suggest varying degrees of morphological
278: relationship between the axonal and dendritic arborizations among all
279: the considered individual cells and also between the three classes
280: BTC, LBC and MC2. In the latter case, the classes LBC e MC2 resulted
281: as being strongly statistically distinct. In brief, the methodology
282: reported in this work has been found to provide a sensitive means for
283: inferring differences between the axonal and dendritic structures.
284: This is a particularly important result, as such differences may be
285: related to important differentiation factors during the neuronal
286: growth, including those caused by genetic modifications (cell
287: predestination), environmental constraints (e.g., interactions with
288: other cells and target specificity), as well as the history of presented
289: stimuli. The identification of such differences also present
290: potential as a complementary subsidy for neuronal cell classification
291: and diagnosis of abnormalities in neurological diseases, which
292: correspond to interesting topics for further research.
293:
294:
295: \begin{acknowledgments}
296:
297: Luciano da F. Costa is grateful to FAPESP (05/00587-5) and CNPq
298: (308231/03-1) for financial support. Marconi S. Barbosa is grateful
299: to FAPESP (02/02504-1,03/02789-9) for sponsoring his post-doc programme.
300:
301: \end{acknowledgments}
302:
303:
304: \bibliographystyle{unsrt} \bibliography{apl}
305:
306: \end{document}
307:
308:
309: __________ NOD32 2115 (20070314) Information __________
310:
311: This message was checked by NOD32 antivirus system.
312: http://www.eset.com
313: