nlin0606053/rp2.tex
1: \newif\ifpdf
2: \ifx\undefined\pdfoutput \pdffalse \else \pdftrue \fi
3: %\documentclass[doublespacing]{elsart}
4: \documentclass{elsart}
5: %\usepackage{changebar,dsfont,tabularx,float,graphicx,amsmath}
6: \usepackage{marvosym,changebar,dsfont,tabularx,graphicx,amsmath}
7: 
8: \ifpdf
9:    \DeclareGraphicsExtensions{.pdf}
10: \else
11:    \DeclareGraphicsExtensions{.eps}
12: \fi
13: 
14: \begin{document}
15: \begin{frontmatter}
16: 
17: \title{Generalised Recurrence Plot Analysis for Spatial Data}
18: \author{Norbert Marwan\corauthref{cor}},
19: \corauth[cor]{Corresponding author.}
20: \author{J\"urgen Kurths}
21: \address{Institute of Physics, University of Potsdam, 14415 Potsdam, Germany}
22: \author{Peter Saparin}
23: \address{Department of Biomaterials, Max Planck Institute of Colloids and Interfaces, 14424 Potsdam-Golm, Germany}
24: 
25: 
26: \begin{abstract}
27: Recurrence plot based methods are highly efficient and
28: widely accepted tools for the
29: investigation of time series or one-dimensional data.
30: We present an extension of the recurrence plots and their 
31: quantifications in order to study recurrent structures in higher-dimensional 
32: spatial data. The capability of this extension is illustrated on 
33: prototypical 2D models. Next, the tested and proved 
34: approach is applied to assess the bone 
35: structure from CT images of human proximal tibia. We find that 
36: the spatial structures in trabecular bone become more self-similar 
37: during the bone loss in osteoporosis. 
38: \end{abstract}
39: \begin{keyword}
40: Data analysis \sep Recurrence plot \sep Spatial data \sep bone structure \sep Osteoporosis
41: \PACS 05.40 \sep 05.45 \sep 07.05.K
42: \end{keyword}
43: \end{frontmatter}
44: 
45: 
46: \section{Introduction}
47: Recurrence is a fundamental property of many dynamical
48: systems and, hence, of various processes
49: in nature. A system may strongly diverge, but after
50: some time it recurs ``infinitely many times as close as one 
51: wishes to its initial state'' \cite{poincare1890}.
52: The investigation of recurrence
53: reveals typical properties of the system and may help to 
54: predict its future behaviour. With the study of nonlinear
55: chaotic systems several methods for the investigation of
56: recurrences have been developed. The method
57: of recurrence plots (RPs) was introduced by Eckmann et~al.~\cite{eckmann87}.
58: Together with different RP quantification approaches
59: \cite{webber94,marwan2002herz}, this method has
60: attracted growing interest for both theory and applications
61: \cite{marwan2003diss}.
62: 
63: Recurrence plot based methods have been succesfully applied
64: to a wide class of data from physiology, geology,
65: physics, finances and others. They are especially suitable for
66: the investigation of rather short and nonstationary data.
67: This approach works with time series or phase space reconstructions
68: (trajectories), i.\,e.~with data which are 
69: at least one-dimensional. 
70: 
71: Recurrences are not restricted to one-dimensional time 
72: series or phase space trajectories. Spatio-temporal processes
73: can also exhibit typical recurrent structures.
74: However, RPs as introduced in \cite{eckmann87} cannot be directly applied to 
75: spatial (higher-dimensional) data. One possible way to study the
76: recurrences of spatial data is to separate the higher-dimensional
77: objects into a large number of one-dimensional data series, and to analyse 
78: them separately \cite{vasconcelos2006}. A more promising approach is
79: to extend the one-dimensional approach of the recurrence plots to
80: a higher-dimensional one.
81: 
82: In the presented work, we focus on the analysis of snapshots of spatio-temporal
83: processes, e.\,g., on static images. An extension of recurrence plots and
84: their quantification to higher-dimensional data is suggested.
85: This extension allows us to apply this method directly to 
86: spatial higher-dimensional data, and, in particular, 
87: to use it for 2D image analysis. We apply this method to 2D pQCT  
88: human bone images in order to investigate differences in 
89: trabecular bone structures at different stages of osteoporosis. 
90: 
91: 
92: 
93: 
94: \section{Recurrence Plots}
95: 
96: The initial purpose of recurrence plots was the visualisation of
97: recurrences of system's states $\vec{x}_i$ in a phase space (with dimension $m$)
98: within a small deviation $\varepsilon$ \cite{eckmann87}. The RP
99: efficiently visualises recurrences even for high dimensional
100: systems. A recurrence of a state at time $i$ at a different
101: time $j$ is marked within a two-dimensional squared matrix
102: with ones and zeros dots (black and white points in the plot), 
103: where both axes represent time. The RP can be formally expressed by the matrix
104: \begin{equation}\label{eq_rp}
105: \mathbf{R}_{i,j} = \Theta\left(\varepsilon-\left\|\vec x_{i} - \vec x_{j}\right\|\right),
106: \quad \vec x_{i} \in \mathds{R}^m,\quad i,j=1 \ldots N,
107: \end{equation}
108: where $N$ is the number of considered states $\vec x_i$, 
109: $\varepsilon$ is a threshold distance (an arbitrary deviation range within 
110: a recurrence is defined), \mbox{$\Vert\cdot\Vert$} denotes
111: a norm and $\Theta (\cdot)$ is the Heaviside function. 
112: 
113: It should be emphasised that this method is a pairwise comparison
114: of system's states at different times along a phase 
115: space trajectory, which is -- although 
116: lying in an $m$-dimensional space -- a one-dimensional curve. The axes of the
117: RP correspond to the time which is given by pursueing a
118: state on the trajectory. Diagonal lines in an RP represent epochs of similar
119: dynamical evolution of the analysed system. For $i=j$ we get
120: the {\it line of identity} (LOI), $\mathbf{R}_{i,i} \equiv 1 \,|_{i=1}^N$, which is
121: the main diagonal line in the RP (Fig.~\ref{fig_rp_smallscalestructures}).
122: 
123: Instead of using the system's states $\vec{x}_i$ which are
124: often unknown, RPs can be created by only using a single time series or a
125: reconstruction of the phase space vectors (e.\,g.~by using
126: time-delay embedding, \cite{takens81}). Such applications 
127: to experimental data have expanded the utilisation of RPs from
128: a tool for the investigation of {\it deterministic phase space dynamics} to a tool
129: for the investigation of {\it similarity and transitions in data series}, 
130: without the rather strong requirement that the data must be from a deterministic
131: dynamical process.
132: The idea of such a similarity plot is not new and can be found
133: in publications earlier than \cite{eckmann87}, e.\,g.
134: in \cite{maizel1981}.
135: This alternative understanding was (unconsciously) the base of the
136: ever increasing amount of application of RPs in data analysis.
137: However, in its present state the RP technique could not
138: be applied on higher-dimensional spatial data.
139: 
140: The initial purpose of RPs was the visual inspection of 
141: the behaviour of phase space trajectories. The appearance of 
142: RPs gives hints about the characteristic time evolution 
143: of these trajectories 
144: \cite{marwan2003diss}. A closer inspection of 
145: RPs reveals small-scale structures which 
146: are {\it single dots}, {\it diagonal lines} as well as {\it vertical}
147: and {\it horizontal lines} (Fig.~\ref{fig_rp_smallscalestructures}). 
148: 
149: \begin{figure}[hbtp] 
150: \centering \includegraphics[width=0.3\columnwidth]{Figs/fig_rp_smallscalestructures} 
151: \caption{Example of a recurrence plot for the logistic map 
152: ($x_{i+1} = ax_i (1 - x)$ with control parameter $a=3.9767$).
153: The RP consists of single dots and line structures.}\label{fig_rp_smallscalestructures}
154: \end{figure}
155: 
156: A {\it diagonal line} 
157: $\mathbf{R}_{i+k, j+k} \equiv 1\,|_{k=0}^{l-1}$
158: (where $l$ is the length of the diagonal line)
159: occurs when one segment of the trajectory runs parallel to another one,
160: i.\,e.~the trajectory re-visits the same region of the phase space at
161: different time intervals. 
162: The length of this diagonal line is determined by 
163: the duration of intervals with similar local behaviour 
164: of the trajectory segments. 
165: We define a line in the RP as a diagonal line of length $l$, if it fulfils the condition
166: \begin{equation}\label{eq_diagonalline}
167:  \left( 1 - \mathbf{R}_{i-1,j-1} \right) 
168:  \left( 1 - \mathbf{R}_{i+l,j+l} \right) 
169: \prod_{k=0}^{l-1} \mathbf{R}_{i+k,j+k}
170:      \equiv 1.
171: \end{equation}
172: 
173: A {\it vertical (horizontal) line} 
174: $\mathbf{R}_{i, j+k} \equiv1 \,|_{k=0}^{v-1}$
175: (where $v$ is the length of the vertical line)
176: marks a time interval in which a system's state does not change 
177: in time or changes very slowly. 
178: It looks like the state is trapped for some time, which is a typical
179: behaviour of laminar states \cite{marwan2002herz}. Because
180: RPs are symmetric about the LOI by definition (\ref{eq_rp}), each vertical
181: line has a corresponding horizontal line. Therefore,
182: only the vertical lines are henceforth considered. Combinations of
183: vertical and horizontal lines form rectangular clusters
184: in an RP.
185: We define a line as a vertical line of length $v$, if it fulfils the condition
186: \begin{equation}\label{eq_verticalline}
187:  \left( 1 - \mathbf{R}_{i,j-1} \right) 
188:  \left( 1 - \mathbf{R}_{i,j+v} \right) 
189: \prod_{k=0}^{v+1} \mathbf{R}_{i,j+k}
190:      \equiv 1.
191: \end{equation}
192: 
193: These small-scale structures are used for the quantitative 
194: analysis of RPs (known as recurrence quantification analysis, RQA). 
195: Using the distributions of the lengths
196: of diagonal lines $P(l)$ or vertical lines $P(v)$, different
197: measures of complexity have been introduced (cf.~\cite{marwan2003diss}
198: for a comprehensive review of definitions and descriptions of 
199: these measures). Here we generalise the measures
200: recurrence rate $RR$, determinism $DET$, averaged diagonal
201: line length $L$, laminarity $LAM$ and trapping time $TT$
202: in order to quantify higher-dimensional data.
203: (cf.~Tab.~\ref{eq_rqa}).
204: 
205: Several measures need a predefined minimal length $l_{min}$ or $v_{min}$,
206: respectively, for the definition of a diagonal or vertical line. These
207: minimal lengths should be as minimal as possible in order to cover 
208: as much variation of the lengths of these lines.
209: On the other hand, $l_{min}$ and $v_{min}$ should be large enough to exclude
210: line-like structures which represent only single, non-recurrent states, which may
211: occur if the threshold $\varepsilon$ is chosen too large or if the data 
212: have been smoothed too stronlgy before computing the RP.
213: 
214: RQA was successfully applied for example for
215: the detection of transitions in event related EEG potentials \cite{marwan2004},
216: the study of interrelations between El Ni\~no and climate in the past \cite{marwan2003climdyn},
217: the investigation of economic data series \cite{gilmore96},
218: of nonlinear processes in electronic devices \cite{elwakil99}
219: or the study of transitions in chemical reactions \cite{rustici99}.
220: For a number of further applications see, e.\,g., \cite{marwan2003diss} or 
221: www.recurrence-plot.tk.
222: 
223: \begin{table}
224: \caption{Generalised recurrence quantification measures 
225: for spatial data of dimension $d$ and with 
226: $\vec \imath,\vec \jmath \in \mathds{Z}^d$.
227: Note that these measures assess recurrence information
228: in terms of length while the original RQA measures
229: quantify it in terms of time.}\label{eq_rqa}
230: \begin{tabularx}{\textwidth}{XXX}
231: \hline
232: RQA measure&equation&meaning\\
233: \hline
234: \hline
235:     recurrence rate	&
236:     $RR =  \frac{1}{N^{2d}} \sum \limits_{\vec \imath,\vec \jmath}^N \mathbf{R}_{\vec \imath,\vec \jmath}$ &
237:     percentage of recurrent states in the system; probability of the recurrence
238:     of any state
239: \\
240:     determinism	&
241:     $ DET_{HS}=\frac{\sum_{l=l_{min}}^N l\, P(l)}{\sum_{\vec \imath,\vec \jmath}^N \mathbf{R}_{\vec \imath,\vec \jmath}}$ &
242:     percentage of recurrence points which form diagonal hyper-surfaces; related to
243:     the predictability of the system
244: \\
245:     laminarity &
246:     $ LAM_{HS}=\frac{\sum_{v=v_{min}}^{N}vP(v)}{\sum_{v=1}^{N}vP(v)}$ &
247:     percentage of recurrence points which form vertical hyper-surfaces; related to
248:     the laminarity of the system
249: \\
250:     averaged diagonal hyper-surface size &
251:     $ L_{HS} = \frac{\sum_{l=l_{min}}^N l\, P(l)}{\sum_{l=l_{min}}^N P(l)}$ &
252:     related to the prediction length of the system
253: \\
254:     trapping size &	
255:     $TT_{HS} = \frac{\sum_{v=v_{min}}^{N} v P(v)} {\sum_{v=v_{min}}^{N} P(v)}$ &
256:     related to the size of the area in which the system does not change
257: \\
258: \hline
259: \end{tabularx}
260: \end{table}
261: 
262: 
263: 
264: 
265: \section{Extension to higher dimensions}
266: Now, we propose an extension of RPs to analyse higher
267: dimensional data. With this step we leave the RPs as a
268: method for investigating deterministic dynamics and
269: focus on its potential in
270: determining similar (recurrent) features in spatial data.
271: 
272: For a $d$-dimensional (Cartesian) system, we define
273: an $n$-dimensional recurrence plot by
274: \begin{equation}
275: \mathbf{R}_{\vec \imath,\vec \jmath} = \Theta\left(\varepsilon-\left\|\vec x_{\vec \imath} - \vec x_{\vec \jmath}\right\|\right),
276: \quad \vec x_{\vec \imath} \in \mathds{R}^m,\ \vec \imath,\,\vec \jmath \in \mathds{Z}^d,
277: \end{equation}
278: where $\vec \imath$ is the $d$-dimensional coordinate vector and $\vec x_{\vec \imath}$
279: is the phase-space vector at the location given by the coordinate vector $\vec \imath$.
280: This means that we decompose the spatial dimension of $\vec x_{\vec \imath}$
281: and consider each space direction separately, e.\,g.~$\vec x_{i_1, i_2, \ldots i_d}$ for
282: $i_1 = 1, \ldots, N$ but $i_2, \ldots, i_d = \text{const}$. Such vectors
283: are now one-dimensional curves in the $m$-dimensional space. 
284: Each of these vectors is pairwisely
285: compared with all others. These individual sub-RPs are the
286: components of the final higher-dimensional RP.
287: The resulting RP has now the dimension $n=2 \times d$ and cannot 
288: be visualised anymore. However, its quantification is still possible.
289: 
290: Similarly to the one-dimensional LOI given by $\mathbf{R}_{i,j} = 1 \ \forall \, i=j$,
291: we can find diagonally oriented, $d$-dimensional 
292: structures in this $n$-dimensional recurrence plot ($n=2\,d$), called
293: the {\it hyper-surface of identity} (HSOI):
294: \begin{equation}
295: \mathbf{R}_{\vec \imath,\vec \jmath} 
296:      \equiv  1 \quad \forall\  \vec \imath = \vec \jmath.
297: \end{equation}
298: 
299: In the special case of a two-dimensional image composed by scalar
300: values, we have 
301: \begin{equation}
302: \mathbf{R}_{i_1,i_2,j_1,j_2} 
303:      \equiv  \Theta\left(\varepsilon-\left\|x_{i_1,i_2} - x_{j_1,j_2}\right\|\right),
304: \end{equation}
305: which is in fact a four-dimensional recurrence plot, and its HSOI 
306: ($\mathbf{R}_{i_1,i_2,i_1,i_2} \equiv 1$) is a two-dimensional plane.
307: 
308: 
309: 
310: 
311: \section{Quantification of Higher-Dimensional RPs}
312: The known RQA is based on the quantification
313: of the line structures in the two-dimensional RPs. Thus, the 
314: definition of higher-dimensional equivalent 
315: structures is crucial for a quantification analysis of
316: higher-dimensional RPs. 
317: 
318: Based on the definition of diagonal lines, Eq.~(\ref{eq_diagonalline}), 
319: we define a diagonal squared hyper-surface of size $\vec l$ ($\vec l = (l,\dots,l),
320: \vec l \in \mathds{Z}^d$) as
321: \begin{equation}\label{eq_diagonalline_x}
322:  \left( 1 - \mathbf{R}_{\vec \imath - \vec 1,\vec \jmath - \vec 1} \right) 
323:  \left( 1 - \mathbf{R}_{\vec \imath + \vec l,\vec \jmath + \vec l} \right) 
324:      \prod_{\substack{k_1, k_2,\dots,\\k_d=0}}^{l-1} \mathbf{R}_{\vec \imath+\vec k,\vec \jmath+\vec k}
325:      \equiv 1.
326: \end{equation}
327: In particular, for the two-dimensional case such a diagonal 
328: hyper-surface (HS) is thus defined as
329: \begin{equation}
330:  \left( 1 - \mathbf{R}_{i_1-1, i_2-1, j_1-1, j_2-1} \right) 
331:  \left( 1 - \mathbf{R}_{i_1+l, i_2+l, j_1+l, j_2+l} \right) 
332:      \prod_{k_1, k_2=0}^{l-1} \mathbf{R}_{i_1+k_1, i_2+k_2, j_1+k_1, j_2+k_2}
333:      \equiv 1.
334: \end{equation}
335: 
336: The next characteristic structure, the vertical squared HS
337: of size $\vec v$ ($\vec v = (v,\dots,v), \vec v \in \mathds{Z}^d$)
338: is defined as
339: \begin{equation}\label{eq_verticalline_x}
340:  \left( 1 - \mathbf{R}_{\vec \imath,\vec \jmath - \vec v} \right) 
341:  \left( 1 - \mathbf{R}_{\vec \imath,\vec \jmath + \vec v} \right) 
342:      \prod_{\substack{k_1, k_2,\dots,\\k_d=0}}^{v-1} \mathbf{R}_{\vec \imath,\vec \jmath+\vec k}
343:      \equiv 1.
344: \end{equation}
345: Its 2D equivalent is
346: \begin{equation}
347:  \left( 1 - \mathbf{R}_{i_1, i_2, j_1-1, j_2-1} \right) 
348:  \left( 1 - \mathbf{R}_{i_1, i_2, j_1+v, j_2+v} \right) 
349:      \prod_{k_1, k_2=0}^{v-1} \mathbf{R}_{i_1, i_2, j_1+k_1, j_2+k_2}
350:      \equiv 1.
351: \end{equation}
352: 
353: 
354: Using these definitions, we can construct the frequency distributions $P(l)$ and $P(v)$
355: of the sizes of diagonal and vertical HS in the 
356: higher-dimensional RP. This way we get generalised RQA measures $DET_{HS}$, $LAM_{HS}$,
357: $L_{HS}$ and $TT_{HS}$ as defined in Tab.~\ref{eq_rqa},
358: which are now suitable for characterising spatial (e.\,g.~two-dimensional) data.
359: 
360: 
361: \begin{figure}[bp] 
362: \centering \includegraphics[width=\columnwidth]{Figs/fig_data_sample} 
363: \caption{Two-dimensional prototypical examples: test 
364: images representing (A) uniformly distributed 
365: white noise, (B) a two-dimensional auto-correlated process (2D-AR2) and (C) 
366: periodical recurrent structures.}\label{fig_data_sample}
367: \end{figure}
368: 
369: \section{Model Examples}
370: 
371: In order to illustrate the ability of the proposed high-dimensional
372: RP's extension, we consider
373: three prototypical model examples from 2D image analysis. The first image (A)
374: is produced by uniformly distributed white noise, the second one (B) is 
375: the result of a two-dimensional auto-correlated process of 2nd order 
376: (2D-AR2; $x_{i, j} = \sum_{k,l=1}^2 a_{k,l}\,x_{i-k,j-l} + \xi$, where $a_{k,l}$ is
377: the 2D matrix of model parameters and $\xi$ is Gaussian white noise) 
378: and the third one (C) represents periodical recurrent 
379: structures (Fig.~\ref{fig_data_sample}). All these example images have a 
380: size of $200 \times 200$~pixels and 
381: are normalised to a mean of zero and a standard
382: deviation of one.
383: 
384: The resulting RPs are four-dimensional matrices of size 
385: $200 \times 200 \times 200 \times 200$, and can hardly be visualised.
386: However, in order to visualise these RPs, we can reduce their dimension by one
387: by considering only those part of the RPs, where 
388: $i_2 = j_2$. The resulting $200 \times 200 \times 200$ cube is a 
389: hypersurface of the four-dimensional RP along the LOI. 
390: For the threshold we use $\varepsilon = 0.2$, which
391: gives clear representations of the RPs.
392: 
393: The features occuring in higher-dimensional RPs provide similar
394: information as known from the classic one-dimensional RPs. 
395: Separated single points correspond
396: to strongly fluctuating, uncorrelated data as it is typical 
397: for, e.\,g., white noise (Fig.~\ref{fig_rp_sample}A).
398: Auto-correlations in data cause extended structures, which can be
399: lines, planes or even cuboids (Fig.~\ref{fig_rp_sample}B).
400: Periodical recurrent patterns in data imply periodic line and
401: plane structures in the RP (Fig.~\ref{fig_rp_sample}C). Two-dimensional
402: slices through such RPs contain similar patterns found by common
403: RPs (Fig.~\ref{fig_rp2_sample}).
404: 
405: 
406: \begin{figure}[tbp] 
407: \centering 
408: \scriptsize{\textsf{A}}\includegraphics[width=0.31\columnwidth]{Figs/rp_noise} 
409: \scriptsize{\textsf{B}}\includegraphics[width=0.31\columnwidth]{Figs/rp_ar2} 
410: \scriptsize{\textsf{C}}\includegraphics[width=0.31\columnwidth]{Figs/rp_period} 
411: \caption{Three-dimensional subsections of four-dimensional RPs of the 
412: images shown in Fig.~\ref{fig_data_sample}. As known from
413: one-dimensional RQA, (A) random data causes homogeneous RPs consisting
414: of single, dis-connected points, (B) correlations in data 
415: cause extended connected structures and
416: (C) periodic data induce periodically occuring structures in the 
417: RPs.}\label{fig_rp_sample}
418: \end{figure}
419: 
420: 
421: \begin{figure}[btp] 
422: \centering 
423: \scriptsize{\textsf{A}}\includegraphics[width=0.31\columnwidth]{Figs/rp_noise_orthoslice} 
424: \scriptsize{\textsf{B}}\includegraphics[width=0.31\columnwidth]{Figs/rp_ar2_orthoslice} 
425: \scriptsize{\textsf{C}}\includegraphics[width=0.31\columnwidth]{Figs/rp_period_orthoslice} 
426: \caption{Slices of the subsections of the four-dimensional RPs 
427: shown in Fig.~\ref{fig_rp_sample}. The similarity to known recurrence 
428: plots is obvious: (A) noise, (B) auto-correlated data and (C) periodic data.}\label{fig_rp2_sample}
429: \end{figure}
430: 
431: 
432: 
433: 
434: We compute the proposed RQA measures (Tab.~\ref{eq_rqa})
435: for the histograms of the sizes of diagonal and vertical 
436: planes (2D HS) in the four-dimensional RPs. For all three examples
437: we use for the minimal size of the diagonal and 
438: vertical HS $l_{min}=3$\,pixels and $v_{min}=4$\,pixels. Although the RQA measures
439: depend on the value of $\varepsilon$, its selection is not crucial for
440: our purpose to discriminate the three different types of structures
441: in the test images. The chosen values
442: for $l_{min}$ and $v_{min}$ are found to be optimal for discriminating
443: the considered images. By choosing smaller values of 
444: $l_{min}$ and $v_{min}$ (but larger than one), the 
445: measures $DET_{HS}$ and $LAM_{HS}$ are closer for 
446: the 2D-AR2 and the periodic image.
447: 
448: \begin{table}
449: \caption{Recurrence quantification measures for the prototypical 
450: examples shown in Fig.~\ref{fig_data_sample}. The measures 
451: are explained in Tab.~\ref{eq_rqa}.}\label{tab_res_examples}
452: \centering \begin{tabular}{lrrrrr}
453: \hline
454: Example&	$RR$&	$DET_{HS}$&	$LAM_{HS}$&	$L_{HS}$&	$TT_{HS}$\\
455: \hline
456: \hline
457: (A) noise	&0.218&	0.007&	0.006&	3.7&	3.0\\
458: (B)	2D-AR2	&0.221&	0.032&	0.065&	3.1&	3.1\\
459: (C)	periodic&0.219&	0.322&	0.312&	5.8&	5.6\\
460: \hline
461: \end{tabular}
462: \end{table}
463: 
464: Four of five RQA
465: measures clearly discriminate between the three types of images 
466: (Tab.~\ref{tab_res_examples}).
467: Only the recurrence rate $RR$ is roughly the same for all test objects. This is
468: because all images were normalised to the same standard deviation.
469: For the random image (A) the determinism $DET_{HS}$ and
470: laminarity $LAM_{HS}$ tend to zero, what is expected, because the values
471: in the image heavily fluctuate even between adjecent pixels. 
472: For the 2D-AR2 image (B), $DET_{HS}$ and $LAM_{HS}$ are slightly above zero, revealing
473: the correlation between adjecent pixels. The last example (C)
474: has, as expected, the highest values in $DET_{HS}$ and $LAM_{HS}$, because
475: same structures occur many times in this image and the image is 
476: rather smooth. Although the trend
477: in $DET_{HS}$ and $LAM_{HS}$ is similar, there is a significant difference
478: between both measures. Whereas $LAM_{HS}$ represents the probability that a
479: specific value will not change over spatial variation (what results
480: in extended same-coloured areas in the image), $DET_{HS}$ measures 
481: the probability that similar changes in the image recur. $LAM_{HS}$ is twice
482: of $DET_{HS}$ for the 2D-AR2 image, obtaining that there are more areas
483: without changes in the image than such with typical, recurrent changes.
484: 
485: 
486: 
487: \section{Application to pQCT data of proximal tibia}
488: 
489: According to the definition of the World Health Organisation, 
490: osteoporosis is a disease characterised by bone loss and changes in the
491: structure of the bone. In the last years, the focus
492: changed to structural assessment of the trabecular bone, because
493: bone densitometry alone is very limited to explain all variation in bone 
494: strength. Furthermore, the rapid progress in the development of new
495: high-resolution computer tomography (CT) scanners 
496: facilitates investigations of the bone micro-architecture. 
497: Different approaches using methods coming from
498: nonlinear dynamics have been recently proposed in order to
499: evaluate structural changes 
500: \cite{benhamou1994,saparin1998,gowin1998,dougherty2001} or 
501: even to predict fracture risks or biomechanical properties
502: \cite{haire1998,majumdar1999,prouteau2004}. These approaches
503: use, e.\,g., scaling properties of bone micro-structure
504: or symbol-encoding of the bone architecture.
505: 
506: Using the RP based method, we will focus here on the  
507: recurrent structures found in images of trabecular bone
508: of proximal tibiae obtained by peripheral 
509: quantitative computer tomography (pQCT). The images were
510: aquired from bone specimens with different stages of 
511: osteoporosis as assessed by 
512: bone mineral density (BMD). Being applied to such images, the RP provides 
513: information about recurrences of bone and soft tissue.
514: 
515: The spatial recurrence analysis is applied to high-resolution
516: pQCT axial slices of human proximal tibia acquired 17~mm below the
517: tibial plateau, with pixel size 
518: 200\,$\mu$m and slice thickness 1\,mm (Fig.~\ref{fig_tibia}). 
519: The images were acquired from 25 bone specimens 
520: with a pQCT scanner XCT-2000 (Stratec GmbH, Germany).
521: The trabecular bone mineral density of these specimens ranges from
522: 30 to 150\,mg/cm$^3$.
523: A standardised image pre-processing procedure was applied to exclude the 
524: cortical shell from the analysis \cite{saparin1998,saparin2002}. The RQA
525: was computed by using the parameters $\varepsilon = 0.04$\,cm$^{-1}$, 
526: $l_{min} = v_{min} = 400$\,$\mu$m. These minimal lengths correspond
527: to two pixels and is found to be appropriate for pQCT images of 
528: such resolution.
529: 
530: In order to further evaluate the proposed RQA measures, we compare them
531: with some recently 
532: introduced structural measures of complexity (SMCs) \cite{saparin1998,saparin2002}. 
533: The SMCs are based on a symbol-encoding of bone elements in
534: the pQCT image. Here we focus on the following SMCs:
535: \begin{enumerate}
536: \item Entropy ($S_a$): quantifies the probability distribution of X-ray attenuation within the ROI;
537: \item Structure Complexity Index (SCI): assesses the complexity and homogeneity of the
538: structure as a whole;
539: \item Trabecular Network Index (TNI) evaluates richness, orderliness, and homogeneity of
540: the trabecular network.
541: \end{enumerate}
542: The computation of the SMCs is applied on the same trabecular
543: area like the RQA measures.
544: 
545: 
546: \begin{figure}[hbtp] 
547: \centering 
548: \includegraphics[width=0.4\columnwidth, angle=180]{Figs/fig_tibia} 
549: \caption{Typical axial pQCT slice of human proximal tibia acquired 17~mm 
550: below the tibial plateau. The trabecular BMD
551: is 65.5\,mg/cm$^3$.}\label{fig_tibia}
552: \end{figure}
553: 
554: The application of the recurrence plot extension to the pQCT images
555: of proximal tibiae reveals a relationship between the recurrences
556: in the trabecular architecture and the osteoporotic stage
557: (Fig.~\ref{fig_rr_bmd} and Tab.~\ref{tab_correlation}). 
558: $RR$ is largest for osteoporotic bone and 
559: shows the strongest relationship
560: with the degree of osteoporosis: it is clearly anti-correlated with BMD (Spearman's 
561: rank order correlation coefficient $R=-0.94$). $DET_{HS}$ and $LAM_{HS}$ are 
562: also maximal for tibiae with high degree of osteoporosis 
563: ($R=-0.66$ and $-0.79$; Fig.~\ref{fig_rqa_bmd}). 
564: We do not find a strong relation between $L_{HS}$, $TT_{HS}$ and BMD.
565: The comparison with the SMCs reveals good relationships 
566: between the RQA measures and $S_a$, SCI and TNI 
567: (Fig.~\ref{fig_rqa_smc} and Tab.~\ref{tab_correlation}).
568: Thus, the RQA measures $RR$, $DET_{HS}$ and $LAM_{HS}$ 
569: contain also information about the
570: complexity and homogeneity of the trabecular network.
571: 
572: 
573: Thus, the proposed RP approach reveals that during the development
574: of osteoporosis the
575: structures in the corresponding pQCT image become more and
576: more recurrent. This is in a good agreement with a decreasing 
577: complexity in the micro-architecture of bone. It confirms
578: the results of an analysis of pQCT images acquired from human proximal
579: tibia and lumbar vertebrae based on symbolic dynamics 
580: \cite{saparin1998,saparin2002}. The direct comparison
581: with the structural quantities (SMCs) shows that the RQA measures
582: provide information about the bone architecture. 
583: The RQA measures reveal a low rate of change 
584: for bone of higher BMD, but higher rate of changes 
585: for specimens with lower BMD (Figs.~\ref{fig_rr_bmd} and \ref{fig_rqa_bmd}).
586: This reflects a higher sensitivity of these measures for
587: osteoporotic trabecular bone and emphasises the nonlinear
588: relationship between the bone architecture as assessed by the
589: RQA measures and bone mass as evaluated by the BMD. As it has been
590: recently shown that the SMCs provide a better estimation of the
591: mechanical bone strength than BMD alone \cite{saparin2005a}, 
592: the proposed RQA measures could further enhance the evaluation to
593: assess the fracture risk of bone.
594: 
595: 
596: \begin{figure}[hbtp] 
597: \centering 
598: \includegraphics[width=.5\columnwidth]{Figs/fig_rr_bmd} 
599: \caption{Recurrence rate $RR$ obtained from
600: four-dimensional RPs of pQCT images of trabecular bone in
601: human proximal tibia of different osteoporotic stages.}\label{fig_rr_bmd}
602: \end{figure}
603: 
604: 
605: 
606: 
607: \begin{figure}[hbtp] 
608: \centering 
609: \includegraphics[width=\columnwidth]{Figs/fig_rqa_bmd} 
610: \caption{Determinism $DET$ (A), laminarity $LAM$ (B),
611: mean line length $L$ (C) and trapping time $TT$ (D) obtained from
612: four-dimensional RPs constructed from pQCT images of trabecular 
613: bone in human proximal tibiae
614: with different degree of osteoporosis.}\label{fig_rqa_bmd}
615: \end{figure}
616: 
617: 
618: 
619: \begin{figure}[hbtp] 
620: \centering 
621: \includegraphics[width=\columnwidth]{Figs/fig_rqa_smc} 
622: \caption{Recurrence rate $RR$ (A) and laminarity $LAM$ (B)
623: compared with trabecular network index TNI and
624: structure complexity index (SCI).}\label{fig_rqa_smc}
625: \end{figure}
626: 
627: 
628: \begin{table}
629: \caption{Rank correlation coefficients $R$ for recurrence
630: quantification measures, BMD and
631: structural measures of complexity (only significant values
632: are shown).}\label{tab_correlation}
633: \centering \begin{tabular}{lrrrrrrrr}
634: \hline
635: 2D-RQA	&BMD		&$S_a$	&SCI	&TNI\\
636: \hline
637: \hline
638: RR		&$-0.94$	&$-0.92$	&$-0.91$&0.84\\
639: DET		&$-0.65$	&$-0.58$	&$-0.61$&0.61\\
640: LAM		&$-0.78$	&$-0.73$	&$-0.75$&0.72\\
641: L		&--			&--			&--		&--	\\
642: TT		&$-0.57$	&$-0.51$	&--		&0.49\\
643: 
644: \hline
645: \end{tabular}
646: \end{table}
647: 
648: 
649: \subsection{Conclusions}
650: 
651: A generalisation of the method of recurrence plots (RPs) and
652: recurrence quantification analysis (RQA) for 
653: the application to higher-dimensional spatial data has been proposed here. 
654: This new method can be used for 2D image analysis,
655: in particular to reveal and quantify 
656: recurrent structures in 2D images.
657: Applying this method on model images, we have shown that it 
658: is able to distinguish typical spatial structures by means of recurrences. 
659: As a first application, we have used this method for the
660: comparison of CT images of human proximal tibia with
661: different degree of osteoporosis. We have found a clear
662: relationship between some of the proposed RQA measures
663: and the complexity and homogeneity of the trabecular structure.
664: Moreover, this approach can be easily extended to higher 
665: dimensions, e.\,g., for 3D analysis of micro-CT
666: images of human bone.
667: This approach will be the base for the further development of 
668: methods for the assessment of structural alteration in trabecular 
669: bone with osteoporosis in patients on Earth
670: or in space flying personnel in microgravity conditions. 
671: 
672: 
673: 
674: \section{Acknowledgments}
675: This study was supported by grants from project MAP AO-99-030 
676: (contract \#14592) of the Microgravity 
677: Application Program/Biotechnology from the Human Spaceflight Program 
678: of the European Space Agency (ESA) and by the European Union
679: through the Network of Excellence BioSim, contract LSHB-CT-2004-005137\&\#65533. 
680: The authors would also like 
681: to acknowledge Scanco Medical AG, Siemens AG, and Roche Pharmaceuticals 
682: for support of the study and thank Wolfgang Gowin and Erika May for preparation
683: and scanning of the bone specimens.
684: 
685: 
686: \clearpage
687: 
688: \bibliographystyle{elsart-num}
689: %\bibliographystyle{plain}
690: \bibliography{mybibs,rp,bone}
691: 
692: \end{document}
693: