q-bio0408004/rato.tex
1: \documentclass{article}
2: 
3: 
4: \usepackage{graphics}
5: \usepackage{graphicx}
6: \usepackage{amsmath}
7: \usepackage{amssymb}
8: \usepackage{subfigure}
9: \begin{document}
10: 
11: \title{Characterizing neuromorphologic alterations with additive shape
12: functionals}
13: 
14: \author{M. S. Barbosa \and L. da F. Costa \and
15:   E. S. Bernardes \and G. Ramakers \and J. van Pelt}
16: 
17: 
18: %% \institute{Cybernetic Vision Research Group, GII-IFSC, Universidade de S\~ ao
19: %%   Paulo, S\~{a}o Carlos, SP, Caixa Postal 369, 13560-970, Brasil \and
20: %%   DFCM-IFSC, Universidade de S\~ ao Paulo, S\~{a}o Carlos, SP, Caixa Postal
21: %%   369, 13560-970, Brasil \and Neurons and Network, Netherlands Institute of
22: %%   Brain Research, Meibergdreef 33, 1105 AZ Amsterdam, The Netherlands}
23: 
24: \maketitle
25: 
26: \begin{abstract}
27:   The complexity of a neuronal cell shape is known to be related to its
28:   function. Specifically, among other indicators, a decreased complexity in
29:   the dendritic trees of cortical pyramidal neurons has been associated with
30:   mental retardation. In this paper we develop a procedure to address the
31:   characterization of morphological changes induced in cultured neurons by
32:   over-expressing a gene involved in mental retardation.  Measures associated
33:   with the multiscale connectivity, an additive image functional, are found to
34:   give a reasonable separation criterion between two categories of cells. One
35:   category consists of a control group and two transfected groups of neurons,
36:   and the other, a class of cat ganglionary cells. The reported framework also
37:   identified a trend towards lower complexity in one of the transfected
38:   groups. Such results establish the suggested measures as an effective
39:   descriptors of cell shape.
40: \end{abstract}
41: 
42: 
43: \section{Introduction}
44: \label{intro}
45: 
46: Classical investigations in neuroscience have added a large amount of
47: descriptive data to the knowledge of the mammalian neural system and
48: ultimately of the functioning of the brain. In recent years the use of
49: computer and algorithmical methods has been incorporated into the biological
50: sciences with the aim of achieving quantitative analysis and modelling. Being
51: no exception to this important tendency, mathematic-computational
52: investigations in neuroscience are still in an incipient stage, with many
53: challenging questions and potential for important results. While the effort to
54: model quantitatively the emergent features of a neural system or element
55: (e.g. a single cell) involves its own technical difficulties, neuroinformatics
56: has already led to important results regarding the interplay between the
57: morphology of dendrites and their firing patterns~\cite{Ooyen:2002}. One
58: difficult aspect implied by the use of computers in neuroanatomy is the
59: relatively limited availability of accurate spatial data of real dendritic
60: arborizations for diverse morphological classes. The alternative approach of
61: generating morphologically realistic neurons has been successfully carried out
62: by using a complementary approach based either on recursive~\cite{Ascoli:1999}
63: or topological~\cite{Pelt:1999,Costa:1999} algorithmic implementation.
64: 
65: The anatomic details of the neuronal cells have been shown to represent a
66: relevant feature, for phylogenic and ontogenic studies, in the diagnosis of
67: diseases and in investigating the interplay between form and function
68: \cite{LU010,LU020,Costa:1999}. With the availability on the Internet of
69: combined databases of virtual and real (histochemically marked) neurons, the
70: analysis or morphological characterization of newly acquired neuronal
71: information can now be the subject of systematic processing by computer
72: methods, provided suitable shape descriptors are identified and applied.
73: 
74: There is a current interest in many branches of Physical and, more recently,
75: Biological sciences to study in a quantitative way the geometrical outline of
76: structures, both static and in development, and from that knowledge to probe
77: further into the functioning and justification of those forms. Examples of
78: results of such a quest in biology ranges from the elucidation of the
79: Fibonacci spirals emerging in
80: sunflowers~\cite{Douady&Coulder:1992,Douady&Coulder:1993} to the fractal
81: characterization of the nature of many botanical plants~\cite{Lind:1990}.  For
82: over a century the search aimed at establishing a relationship between the
83: form and function of neuronal cells has challenged and stimulated researchers,
84: starting from the pioneering efforts of Ram{\'o}n y
85: Cajal~\cite{Cayal:1894}. While many diseases have been diagnosed based on the
86: visual perception of morphological alterations in general tissue morphology,
87: automated approaches remain promising subjects of research, with many
88: potentially important applications.
89: 
90: Although it is now generally accepted that dendritic morphology plays a
91: crucial role in the functioning of the neural cell and ultimately in the
92: behaviour of the neural
93: system~\cite{Ooyen:2002,Washington:2000,Uylings:2002,Devaud:2000,Brain:2003},
94: there is no deterministic way to choose the best quantitative descriptor of
95: the geometry or topology of single neurons.  In order to be particularly
96: useful, such descriptors should correlate in some way with the respective
97: performed functions and behaviour, allowing abnormalities and specific
98: divergences from a healthy state or development to be clearly identified. An
99: additional benefit of such an investigation of the physical basis of neuronal
100: shape and growth is the possibility for investigating, through modelling, the
101: relationship between neuronal shape and function.
102: 
103: The study of biological forms requires the selection of shape
104: descriptors or shape functionals that fulfil a set of
105: requests. First, measures are expected to be objective and fast (in
106: order to treat a representative number of cases), while retaining
107: comprehensiveness and being potentially discriminative.  In addition,
108: such measures should lead to meaningful biological interpretations. An
109: important theoretical issue is the relative degeneracy of a set of
110: measures, a concept that reflects the fact that many different forms
111: or shapes produces the same or approximated measures. It has been
112: experimentally shown~\cite{LU010,LU021,LU099} that multiscale analysis
113: tends to augment the resolution power of geometric descriptors and,
114: despite the increased computational demands, this procedure provides
115: useful additional information and characterization of the considered
116: objects.  The neuronal cells have, in general, a complex spatial
117: tree-like structure with many dendritic bifurcations leading to
118: efficient spacial coverage, amplifying the influence area and
119: optimally connecting the neuron with a neighbouring cell's dendrites or
120: axons. This structure has been shown to exhibit different levels of
121: partial fractality at different spatial
122: scales~\cite{MOR89,Coelho:1996,LU020}.
123: 
124: A recently reported~\cite{Raedt01,Raedt02} procedure for calculating
125: additive shape functionals, known as Minkowski functionals gathered in
126: a framework called Integral-Geometry Morphological Image Analysis,
127: MIA, has been successfully applied to many areas of research including
128: Statistical Physics, Cosmology and Material
129: Sciences~\cite{Santalo,Stoyan,Raedt01,Raedt02,Mecke}. A first attempt
130: to bring those morphological concepts into biological sciences,
131: targeting the efficient discrimination of two main types of the
132: domestic cat's ganglionary neuronal cells, has been reported
133: recently~\cite{Barbosa:2002}. The use of additive functionals is
134: implemented in a pixelwise approach leading to the possibility of
135: straightforwardly obtaining the above mentioned multiscale fractal
136: dimension for complementary shape description. Every additive
137: continuous and motion invariant functionals in the 2D Euclidean plane
138: can be expressed as a combination of three Minkowski functionals which
139: are proportional to a known geometric quantity, namely the metric area
140: and perimeter and the topological Euler connectivity number. This
141: completeness extends to higher dimensions and the geometrical
142: functionals multiplicity is always one plus the dimension of the
143: lattice generalized voxels (e.g 3 in 2D plane, 4 in 3D space, etc.).
144: While the theory of Integral geometry provides a sophisticated set of
145: results and formulae, the practical implementation of the procedures
146: in Image analysis is relatively simple and
147: efficient~\cite{Raedt01,Raedt02}. The basic ideas of this methodology
148: is described in section~\ref{AF} and a simple example is worked out to
149: outline the procedure.
150: 
151: In this paper we investigate the importance of the spatial
152: distribution of branching points in neuronal arborizations to provide
153: a discriminative and informative characterization of the neuronal
154: morphology, with special attention given to additive shape
155: functionals. This novel procedure involves mapping the neuronal image
156: onto a set of points representing its bifurcation pattern, which are
157: subsequently dilated in order to produce a multiscale
158: representation~\cite{TAD01}, as the Minkowski functionals of such a
159: set of points is recorded. The obtained results substantiate the
160: potential of the above procedure regarding its application to a
161: database of 3 categories of rat neuronal cells and one class of cat
162: ganglion cells. The results are in accordance with the biological
163: importance of the spatial distribution of
164: branches~\cite{MOR89,SHO53,KRAPP98,Ooyen:2002,Uylings:2002,Washington:2000}
165: and provide a discriminative measure for the morphological
166: characterization of the neurons, indicating subtle morphological
167: differences as a consequence of the considered gene transfectation in
168: the rat categories.  This suggests that those treatments may not
169: significantly alter the topological aspect of the branching cell form
170: as far as the considered measurements are concerned.
171: 
172: \section{Additive Functionals}\label{AF} 
173: 
174: Integral geometry algorithms have been successfully used to
175: characterize morphologically complex patterns where the precise
176: process of formation is not completely known and is a subject of
177: modelling, see for example~\cite{Mecke:1996}. The central procedure is
178: the calculation of intrinsic volumes or Minkowski functionals (or yet
179: {\it querrmassintegrals}), which are a generalization of the usual
180: determination of volume. They can be defined as related to both
181: differential and integral geometry setups.  
182: 
183: \begin{figure}[htb]
184: \begin{centering} \includegraphics[scale=.25,angle=-90]{oito.ps}
185: \caption{The disjoint decomposition of a pattern $K$ (left) into a sum
186:   $\mathcal{P}$ (right) of disjoint interior elements. \label{fig:1}}
187: \end{centering}
188: \end{figure}
189: 
190: For the sake of a better explanation, we start with a description of
191: the practical aspect of the adopted procedure. The Minkowski
192: functionals of a body $K$ in the plane are proportional to its area
193: $A(K)$, perimeter $U(K)$ and the connectivity or Euler number
194: $\chi(K)$. The usual definition of the connectivity from algebraic
195: topology in two dimensions is the difference between the number of
196: connected $n_c $ components and the number of holes $n_h$,
197: 
198: \begin{equation} 
199: \chi(K)=n_c-n_h.  
200: \end{equation} 
201: 
202: So, if $K$ happens to be an image of the number 8 as in
203: Figure~\ref{fig:1}, its connectivity number would be $\chi(K
204: )=1-2=-1$. One instance where these functionals appear naturally is
205: while attempting to describe the change in area as the domain $K$, now
206: assumed to be convex, undergoes a dilation through a parallel set
207: process using a ball $B_r$ of radius r
208: 
209: \begin{equation}
210: \label{eq:dila}
211: A(K \oplus B_r)=A(K)+U(K)r+\pi r^2. 
212: \end{equation}   
213: 
214: Generalizing to higher dimensions, the change in hyper volume is given
215: by the Steiner formula~\cite{Santalo}
216: 
217: \begin{equation}
218: \label{eq_dilageral}
219: v^d(K\oplus B_r)=\sum_{\nu=0}^d \binom{d}{\nu}
220: W_{\nu}^{(d)}(K)r^{\nu}, 
221: \end{equation} 
222: 
223: where the coefficients $W_{\nu}^{(d)}$ are referred to as Minkowski
224: functionals. For instance in the plane ($d=2$),
225: 
226: \begin{equation}
227: \label{eq:coeffs}
228:   W_0^{(2)}(K)=A(K), W_1^{(2)}(K)=\frac{U(K)}{2}, W_2^{(2)}(K)=\chi(K)\pi.
229: \end{equation}
230: 
231: Despite the wealth of results and continuum formulae for obtaining
232: these functionals, it is useful to recur to the discrete nature of the
233: binary images we wish to analyse by looking at the distribution of
234: pixels in the planar lattice. On that level, by exploring the
235: additivity of the Minkowski functionals, their determination reduces to
236: counting the multiplicity of basic building blocks that compose the
237: object in a disjointed fashion. Figure \ref{fig:1} illustrates this
238: process. A pixel can be decomposed as a disjointed set of 4 vertices, 4
239: edges and one square. The same process can be applied to any object in
240: a lattice. The fundamental information needed here is a relationship
241: for the functionals of an open interior of a $n$-dimensional body $K$
242: which is embedded into a $d$-dimensional space
243: 
244: \begin{equation}\label{eq_open} 
245: W_{\nu}^{(d)}(\breve{K})=(-1)^{d+n+\nu}
246: W_{\nu}^{(d)}(K), \nu=0,\ldots, d.
247: \end{equation} 
248: 
249: As there is no overlap between these building blocks and by the
250: property of additivity of these functionals, for a body $P$ composed
251: of disjointed convex interior pieces $\breve{N}_m$, we may write
252: 
253: \begin{equation}\label{eq_whole} 
254: W_{\nu}^{(d)}(\mathcal{P})=\sum_{m}
255: W_{\nu}^{(d)}(\breve{N}_m)n_m(\mathcal{P}), \; \nu=0,\ldots,d.
256: \end{equation} 
257: 
258: Where $n_m(\mathcal{P})$ stands for the number of building elements of
259: each type $m$ occurring in the pattern $\mathcal{P}$. For a
260: two-dimensional space, which is our interest regarding the considered
261: neuronal images, we display in Table~\ref{tb_buildblocks} the value of
262: Minkowski functionals for the building elements in a square lattice of
263: pixels and their direct relation to familiar geometric quantities on
264: the plane. Using the information (with $a=1$) presented in
265: Table~\ref{tb_buildblocks} and equation \eqref{eq_whole} we have
266: 
267: \begin{equation}\label{eq_func} \quad
268: A(\mathcal{P})=n_2,\;U(\mathcal{P})=-4n_2+2n_1,\;\chi(\mathcal{P})=n_2-n_1+n_0.
269: \end{equation} 
270: 
271: Going back to Figure~\ref{fig:1}, we find, for this specific example,
272: as $n_2=16$, $n_1=47$ and $n_0=30$, that $A=16$, $U=30$ and $X=-1$.
273: So the procedure to calculate Minkowski functionals of a pattern $K$
274: has been reduced to the proper counting of the number of elementary
275: bodies of each type that compose a pixel (squares, edges and vertices)
276: involved in the make up of $\mathcal{P}$.  
277: 
278: \begin{table}[htb]
279: \caption{Minkowski functionals of open bodies $\breve{N}_m$ which
280: compose a
281:   pixel $K$: $\breve{P}$ ({\it vertex } ), $\breve{L}$ ({\it open edge}) and $\breve{Q}$ ({\it open square}). \label{tb_buildblocks}}
282: \begin{centering}
283: \begin{tabular}{l|l|l|l|l}\hline\hline
284: $m$ & $\breve{N}_m$  & $W_0^{(2)}=A(\breve{N}_m)$ & $W_1^{(2)}=\frac{1}{2}
285: U(\breve{N}_m)$ & $W_2^{(2)}=\pi\chi(\breve{N}_m)$ \\ \hline 
286: 0 &  $\breve{P}$   & 0           & 0           & $\pi$    \\ \hline
287: 1 &  $\breve{L}$   & 0           & $a$         & $-\pi$   \\ \hline
288: 2 &  $\breve{Q}$   & $a^2$       & $-2a$       & $\pi$    \\ \hline\hline
289: \end{tabular}
290: \end{centering}
291: \end{table}
292: 
293: 
294: \section{Branching point patterns} The existence of profuse branching
295: structures in natural shapes provides evidence of the effectiveness of
296: such shapes in providing an interface with its environment as well as for
297: enabling connections. As pointed out in~\cite{Uylings:2002} many
298: conditions may interfere with the morphology of the neuronal tree
299: structure, for instance learning, 'enriched' environment, hormonal
300: fluctuations and levels of bioelectric activity.
301: 
302: The relationship between dendritic morphology and cell functioning,
303: alone or connected in networks, has drawn the attention of scientists
304: leading to a search for a set of measures that would as completely as
305: possible describe the neuronal shape~\cite{Hillman:1979}. For example,
306: the dependence of the dendritic diameter on the distance from soma,
307: the relationship between the dendritic diameter before a bifurcation
308: point and the diameter of the two daughters stemming from this point,
309: or the ratio of diameters of daughter dendrites. These are standard
310: examples of a global feature (the first) and two local properties (the
311: following two) one can specify/calculate in describing/analysing a
312: neuronal shape, real or virtual, see~\cite{Ascoli:1999}. There is a
313: vast recent literature describing shape analysis in general and
314: applied specifically to neuronal shape, see for
315: example~\cite{Uylings:2002,Kossel,KRAPP98} for a general dendritic
316: description, and~\cite{Coelho:1996} for multiscale fractality,
317: \cite{LU021} for wavelets and ~\cite{LU010} for bending energy applied
318: to neuromorphometry.
319: 
320: 
321: Since the pioneering work of Sholl~\cite{SHO53}, some authors have
322: mentioned specific situations where the well-known Sholl analysis
323: could lead to degenerate results, assigning to visually different
324: neurons the same (in statistical terms) descriptor values,
325: see~\cite{Kossel} for example. The method makes use of a reference
326: point at the soma and draws concentric circles through the dendritic
327: field, while counting the number of intersections within each
328: circle. While this process is easy to implement manually, which may
329: account for its popularity, it clearly fails for asymmetric
330: forms. While alternative approaches~\cite{Uylings:1989b} report
331: success in improving this setup, we take a different route, abandoning
332: the soma as a reference point.
333: 
334: We start our analysis by taking all bifurcation points and
335: constructing another image keeping the metric relation between these
336: points. Although there are automatic ways of extracting salient
337: points~\cite{LU078,LU009,costa:1997}, we implemented this task in a
338: semi-automated way, which is followed by a sequence of exact dilations
339: (permitted parallel set dilation~\cite{TAD01}) of the selected
340: points. Although primarily interested in the connectivity, all
341: Minkowski functionals are calculated by a routine that counts
342: efficiently the multiplicity of the building elements for each neuron
343: image at every radius of dilation as described in section~\ref{AF},
344: equation~\ref{eq_func}. The idea is to use the connectivity of the set
345: of points to capture, at each scale, the spatial relationship of the
346: neuron branching structure. As this measure is invariant to rotation
347: and translation, as well as to scaling, eventual variations of size
348: due to the relative stage of development can be disregarded.
349: 
350: We investigate the morphology of two different categories of cells. One
351: involves cat ganglion alpha cells, which constitutes a class by itself
352: but rather diversified in form,
353: see~\cite{Coelho:2002,Barbosa:2002}. The other includes foetal rat
354: cerebral cortex neurons cultured in a 2D tissue
355: system~\cite{Ramakers:1998}. This category is subdivided into three
356: differently treated classes consisting of a control group, a group of
357: neurons transfected with a gene OPHN1 which encodes for oligo-phrenin,
358: and a positive control group transfected with p190 RhoGAP.
359:             
360: Figure~\ref{fig:2} shows examples of two treated rat neuronal cells,
361: illustrating the branching selection process adopted in this
362: work. Note that the neuron images have been rescaled in this picture,
363: but the final dilation radius in all processed samples are the same
364: and equal to 30 pixels.
365: 
366: \begin{figure*}[p] 
367: \begin{centering}
368: \begin{tabular}{lll}
369: \subfigure[OPHN1]{\includegraphics[scale=.3,angle=  0]{neuron.ps}} &
370: \subfigure[]{\includegraphics[scale=.3,angle=0]{points1.ps}}&
371: \subfigure[]{\includegraphics[scale=.3,angle=0]{rato_fim.ps}}\\
372: \subfigure[RhoGAP]{\includegraphics[scale=.25,angle=  0]{neuron2.ps}}&
373: \subfigure[]{\includegraphics[scale=.25,angle=0]{points2.ps}}&
374: \subfigure[]{\includegraphics[scale=.25,angle=0]{Rho12.ps}}\\
375: \multicolumn{3}{c}{
376: \subfigure[]{\includegraphics[scale=1.6,angle= -90]{measures.ps}}}
377: \end{tabular}
378: \caption{Isolating morphological information relating to the neuronal branching
379:   structure while preserving the metric relationship among the branching
380:   points. Each bifurcating point undergoes an exact dilation while its
381:   connectivity is recorded as function of the scale parameter. 
382:   Prototypical transfected rat neuronal cell, a) and d). Extracted pattern of
383:   bifurcation points, b) and e). The end of the bifurcation point parallel set
384:   dilation procedure, c) and f). The resulting discriminating functional for
385:   both neuron types, g).\label{fig:2}}
386: \end{centering}
387: \end{figure*}
388: 
389: \begin{figure*}[p]
390: \begin{centering}
391: \includegraphics[scale=0.45,angle=-90]{ratoX_scatter.ps}% 
392: \caption{A discriminative feature space. Most of the Rho-Gap transfected
393: neurons went to the lower left corner of the plot, while the Alpha cat cells
394: are spread by themselves towards higher values. There is a fuzzy display of both
395: control and OPNH-1 samples. \label{fig:3}}
396: \end{centering}
397: \end{figure*} 
398: 
399: \begin{figure*}[p]
400: \begin{centering}
401: \includegraphics[scale=0.6,angle=0]{symbol.eps}% 
402: \caption{The discriminative power of the above feature space in depth. A
403:   discriminant analysis reveals a reasonable separation among ALPHA and Rho-GAP
404:   groups and is not conclusive for the others.\label{fig:4}}
405: \end{centering}
406: \end{figure*} 
407: 
408: It is important to comment on the complementary nature of the
409: framework proposed in this article with respect to more traditional
410: approaches reported in the literature.  When compared to the previous
411: Sholl methodologies, the measurements in our approach adopt
412: multiscale concepts so as to retain as much as possible the metric
413: and global structure, while avoiding the somewhat arbitrary use of the
414: soma as a general reference.  In order to express in a more objective
415: and effective way the morphological features of special interest,
416: namely the overall complexity of the neuronal pattern, our procedure
417: was designed from the start to concentrate on the bifurcation points,
418: which are used as primary data.  No attempt is made here to achieve
419: completeness in the shape description, or to use the considered
420: measures as features to be incorporated in an algorithm to grow
421: artificial neurons.  However, the reported measurements can be useful
422: for statistical validation of such artificial data, in the sense that
423: both the simulated and original neuronal structures should lead to the
424: same probability densities.
425:    
426: 
427: \section{Results} 
428: For each cell we calculate multiscale curves of connectivity as shown in
429: Figure~\ref{fig:2}. As discriminating measurements we take the area under the
430: interpolating curve (its integral) and the standard deviation of the
431: difference between that curve and the original data points. The former can
432: provide information about the overall structure, while the latter
433: characterized the finer details. Having performed similar analysis for area
434: and perimeter functionals, as well as for the multiscale fractality as derived
435: from the area data, we decided to focus on the multiscale connectivity of the
436: branching points set as a descriptor for its relationship to the spatial
437: distribution of branching points and its above average discriminating power.
438: 
439: Our primary concern here is to investigate the effectiveness of the
440: procedure adopted to quantify the subtle morphological aspects that
441: may characterize different cell types while also expressing a
442: biologically relevant attribution of the cell form, namely the
443: distribution of bifurcation points. In this regard, our results, shown
444: by Figure~\ref{fig:3}, led to a well-defined separation of the alpha
445: cells in a feature space defined by the integral of the multi scale
446: connectivity and its standard deviation.
447: 
448: Although the three differently treated rat cells are not clearly
449: distinguishable, one group of them, namely the one containing neurons
450: which show over expression of RHO-GAP following transfectation, clearly
451: tends to populate the lower left corner of the feature space. Yet,
452: while high dispersion is shown by the rats sample cells, including
453: the control group which is as much diversified in form as the treated
454: groups, the statistical tendency towards complexity agrees with a
455: previous analysis which focused on the multiscale fractal dimension of
456: this group of cells\cite{LU101}.
457: 
458: Figure~\ref{fig:4} shows a scatter plot defined by two canonical
459: discriminant functions (their canonical coefficients are displayed in
460: Table~\ref{tab:CC}) for the above selected shape descriptors. The
461: separability in groups is optimized (see Table~\ref{tab:MD} for a
462: quantitative index, the Mahalanobis distance) in such an analysis,
463: leading to a somewhat improved visibility of the above mentioned
464: distinction of the Rho-GAP over expressing set of cells from the
465: others. Nonetheless the other differently treated group mixes up
466: strongly with the control group, as can be verified from
467: Table~\ref{tab:Plug} which presents the plug-in (also called
468: confusion matrix) classification results. As can be seen from this
469: statistical analysis the alpha cat cell is not trivially
470: distinguishable from the others, as shown by the overlaps in both the
471: feature and canonical discriminant scatter plots. While the
472: effectiveness of use of Minkowski functionals (considering the whole
473: cell) has been previously reported~\cite{Barbosa:2002}, the framework
474: developed here is meant to emphasize the distribution of the
475: bifurcation points, in a way similar to that in Sholl's analysis, on
476: the cell morphology. 
477: 
478: \begin{table}
479: \caption{Canonical Coefficients}
480: %\begin{centering}
481: \begin{tabular}{l|l|l}\hline\hline
482:    &        dim1       & dim2\\ \hline
483: Connectivity Integral &-0.893114489  &1.08447785\\ \hline
484: Connectivity Standard Deviation  &-0.003344825 &-0.00731277\\ \hline\hline
485: \end{tabular}
486: \label{tab:CC}
487: %\end{centering}
488: \end{table}
489: 
490: \begin{table}
491: \caption{Mahalanobis Distance}
492: \begin{centering}
493: \begin{tabular}{l|l|l|l|l}\hline\hline
494:         &   Alpha &     CMV   & OPHN1  &Rho-GAP\\ \hline
495:   Alpha &0.00 &6.08 &6.53 &8.97\\ \hline
496:     CMV &         &0.00 &0.10 &0.27\\ \hline
497:   OPHN1 &         &         &0.00 &0.31\\ \hline
498: Rho-GAP &         &         &         &0.00\\ \hline\hline
499: \end{tabular}
500: \label{tab:MD}
501: \end{centering}
502: \end{table}
503: 
504: 
505: \begin{table*}
506: \caption{Plug-in classification table}
507: \begin{centering}
508: \begin{tabular}{l|l|l|l|l|l|l}\hline\hline
509:         &Alpha &CMV &OPHN1 &Rho-GAP &    Error &Posterior.Error\\ \hline
510:   Alpha &   20 &  1 &    3 &      1 &0.20 &      0.25\\ \hline
511:     CMV &    0 &  9 &   10 &     11 &0.70 &      0.73\\ \hline
512:   OPHN1 &    2 &  6 &   10 &     13 &0.67 &      0.65\\ \hline
513: Rho-GAP &    1 &  6 &    5 &     19 &0.38 &      0.41\\ \hline\hline
514: \end{tabular}
515: \label{tab:Plug}
516: \end{centering}
517: \end{table*}
518: 
519: 
520: \section{Conclusions} 
521: This work describes how promising results have been obtained regarding a novel
522: procedure and measurements for classifying different types of neurons and to
523: reveal morphologically relevant attributes of different classes.  Although the
524: subtle morphological variations induced by the treatment of our samples are
525: masked by the statistical distributions and show only a small tendency towards
526: a reduction in complexity, the procedure shows clearly its discriminating
527: power when applied to a well characterized class of alpha cat ganglionary
528: cells. As the considered measurements were designed to express a particular
529: biological aspect of the neuronal geometry, namely the spatial relationship
530: among the branching points, we would expect that this specific trait is not
531: significantly affected by the gene transfectation process.  The procedure
532: reported in this paper can be generalised to process 3D Neurons, real or
533: virtual, and can be implemented straightforwardly. By extending the present
534: analysis to 3D we suspect that subtle information relating to neuron
535: complexity could be significantly improved.
536:  
537: 
538:   
539: 
540: \section{Acknowledgements}
541: 
542: The authors thank Regina C{\'e}lia Coelho for kindly providing of the
543: Cat Neurons image database. This work was financially supported by
544: FAPESP (processes 02/02504-01, 99/12765-2 and 96/05497-3) and to CNPQ
545: (process 301422/92-3).
546: 
547: 
548: 
549: 
550: \clearpage            
551: %\nocite{*}
552: \bibliographystyle{unsrt}
553: \bibliography{rato}
554: 
555: 
556: 
557: 
558: 
559: 
560: 
561: \end{document}
562: 
563: 
564: 
565: 
566: 
567: