1: \documentclass[aps,prl,twocolumn,epsf,psfig,superscriptaddress]{revtex4}

2: \usepackage[dvips]{graphicx}

3: \draft

4: \begin{document}

5: \newcommand{\bea}{\begin{equation}}

6: \newcommand{\ber}{\begin{eqnarray}}

7: \newcommand{\eea}{\end{equation}}

8: \newcommand{\eer}{\end{eqnarray}}

9: \newcommand{\ct}{\cite}

10: \newcommand{\bi}{\bibitem}

11: %\baselineskip = 14.0 true pt

12: \title

13: {Universality in the merging dynamics of parametric active contours: a study in

14: MRI-based lung segmentation}

15: %

16: \author{Amit K. Chattopadhyay}

17: \affiliation{

18: Max Planck Institute for the Physics of Complex Systems,

19: Noethnitzer Strasse 38,

20: 01187 Dresden, Germany

21: }

22: 

23: \email{akc@mpipks-dresden.mpg.de}

24: 

25: %

26: 

27: \author{Nilanjan Ray}

28: \affiliation{

29: Department of Electrical and Computer Engineering, 

30: Thornton Hall, 

31: 351 McCormick Road, 

32: P.O. Box 400743,  

33: Charlottesville, VA 22904-4743, U.S.A.} 

34: \email{nray@virginia.edu, acton@virginia.edu}

35: %

36: \author{Scott T. Acton}

37: 

38: \affiliation{

39: Department of Electrical and Computer Engineering, 

40: Thornton Hall, 

41: 351 McCormick Road, 

42: P.O. Box 400743,  

43: Charlottesville, VA 22904-4743, U.S.A.}

44: \email{acton@virginia.edu} 

45: %

46: \date{\today}

47: %

48: \iffalse

49: \footnote{emails: akc@mpipks-dresden.mpg.de \\

50: nray@virginia.edu \\

51: acton@virginia.edu}

52: \fi

53: 

54: %

55: 

56: \begin{abstract}

57: \noindent

58: Measurement of lung ventilation is one of the most reliable 

59: techniques of diagnosing pulmonary

60: diseases. The time consuming and bias prone traditional methods using 

61: hyperpolarized H${}^{3}$He and ${}^{1}$H magnetic resonance imageries have 

62: recently been improved by an automated technique based on multiple active 

63: contour evolution. Mapping results from an equivalent thermodynamic model, 

64: here we analyse the fundamental dynamics orchestrating the active contour (AC)

65: method. We show that the numerical method is inherently connected 

66: to the universal scaling behavior of a classical nucleation-like dynamics. 

67: The favorable comparison of the exponent values with the theoretical 

68: model render further credentials to our claim. \\

69: 

70: \noindent

71: PACS: 87.10. +e, 87.68. +z, 82.60. Nh

72: \end{abstract}

73: 

74: \pacs{87.10. +e, 87.68. +z, 82.60. Nh}

75: %

76: 

77: \maketitle

78: 

79: 

80: 

81: \noindent

82: 

83: 

84: 

85: Ventilation analysis is an authentic way of diagnosing lung airway diseases.

86: The ratio of the volume of ventilated (functional) portions of lungs

87: to the total lung volume is known as lung ventilation, which is

88: used in validating pulmonary drugs \cite{Ray}. The process involves 

89: two complementary 

90: magnetic resonance (MR) imaging modalities, the hyperpolarized 

91: helium-3 ($\rm{H}^{3}$He) imagery and the proton (${}^{1}$H) imagery. Lung 

92: functionality, including the volume of ventilated lungs, can be obtained from 

93: the former modality while lung anatomic details, including total lung volume, 

94: are accessed through the latter \cite{Ray}. Since manual investigation of

95: the MR 

96: imagery to compute lung ventilation is extremely time consuming,  

97: an {\em active contour (AC)} or {\em snake} based 

98: automated method has been proposed \cite{Ray,Xu} to compute the total 

99: lung volume from proton MR

100: imagery on a 

101: 2-D slice-by-slice basis \cite{Ray}. A snake is defined as a massless 2-D thin

102: string 

103: (closed or open) that can move on the image domain driven by  

104: two types of forces, – internal elastic forces and external image

105: forces 

106: \cite{Kass}. Under the influence of these two forces an initial

107: contour (snake) clings to 

108: image edges and delineates an object. A snake always gives continuous

109: edges unlike any traditional edge detector (such as the Canny method 

110: \cite{Canny}), thereby eliminating any post-processing steps to

111: connect the 

112: detected broken edges. These two properties are particularly useful

113: when the 

114: object outline is broken and noisy as in most of the ${}^{1}$H MR

115: imagery. 

116: The snake method of finding the object boundary relies on the initial 

117: snake placement inside the image. If a small initial snake is placed 

118: inside a lung cavity on the MR image, while growing, the snake may be 

119: stopped by the associated numerical artifacts and may not capture the 

120: actual lung outline 

121: \cite{Ray}. Starting with a larger snake may result in missing the lung

122: cavity completely. A possible solution is to start

123: with multiple non-overlapping small snakes inside the lung cavity

124: and evolve (grow) them until they merge with each other and capture 

125: the cavity outline \cite{Ray}. During such a process, the growing 

126: snakes merge with each other into a single contour. This automatic

127: merging of non-overlapping snakes is characterized by certain attributes:

128: a) during the evolution process no 

129: two snakes overlap with each other, b) every snake stops evolving at

130: the object edge as a single snake does during its course of evolution,

131: and c) growing convex shaped snakes (e.g., circular or rectangular 

132: snakes) inside a convex object recovers the object boundary. 

133: Although merging of multiple 

134: snakes is experimentally verified in a multitude of cases,  

135: a concrete theoretical understanding of 

136: this merging snake approach remains a challenge. 

137: 

138: The aim of this letter 

139: is to bridge this gap by showing that the underlying principle of 

140: multiple snake-merging is governed by a {\em universal power law} behavior 

141: which originates from an inherent {\em nucleating} structure.  

142: From a biomedical engineering perspective, the study of the efficacy of the

143: lung cavity delineation method is crucial for robust clinical application. 

144: The lung cavity segmentation by merging snake method 

145: involves three steps: a) initially small 

146: non-overlapping contours are placed inside the lung cavity, b)

147: generalized gradient vector flow (GGVF) fields \cite{Xu} are

148: computed with a 

149: Dirichlet boundary condition on the initial circular snakes 

150: \cite{Ray}, and c) all the snakes are evolved simultaneously and 

151: independently of each other with the GGVF force field as the 

152: external force for the snakes. This automated lung cavity 

153: segmentation is attractive for a number of reasons. While 

154: other merging snake algorithms, such as the one proposed by 

155: McInerney and Terzopoulos \cite{McInerney} is computationally 

156: non-trivial compared to the original snake evolution algorithm 

157: of Kass et al. \cite{Kass}, the merging snake algorithm by 

158: Ray {\em et al}. maintains the same computational simplicity of 

159: Kass {\em et al}. algorithm. Also, based on the position 

160: of an initial snake, the rigidity parameters of the snakes can be

161: varied, so that on one hand delicate high curvature features, such as 

162: costophrenic angles, can be accurately captured, 

163: while on the other hand, snakes can be made sufficiently stiff in order 

164: to avoid capturing artifacts. Fig. 1 shows multiple snake 

165: initialization, evolution, merging and delineation of lung boundary 

166: by the Ray {\em et al’l} method.

167: 

168: %

169: 

170: \begin{figure}

171: %\includegraphics[height=4.0cm,width=8.0cm,angle=0]{/home/akc/nilanjan/paper/fig1.ps}

172: \includegraphics[height=4.0cm,width=8.0cm,angle=0]{fig1.ps}

173: \caption{Evolution of multiple snakes in a 2-D slice of the lungs by the Ray,

174: {\em et al} \cite{Ray} method.} 

175: \label{Fig:1}

176: \end{figure}

177: 

178: %

179: 

180: \begin{figure}

181: 

182: %\includegraphics[height=4.0cm,width=8.0cm,angle=0]{/home/akc/nilanjan/paper/fig2.ps}

183: \includegraphics[height=4.0cm,width=8.0cm,angle=0]{fig2.ps}

184: \caption{Evolution of snakes (here circular, although one can use 

185: rectangular snakes as well) in a hypothetical circle image.}

186: \label{Fig:2}

187: 

188: \end{figure}

189: 

190: %

191: 

192: In order to map the merging snake scenario to statistical 

193: thermodynamical systems, we first provide the mathematical background 

194: for a parametric active contour or snake. A snake is a curve 

195: ${\bf C}(s) = (p(s),q(s))$

196: defined by the parameter $s \in [0,1]$. The snake is evolved

197: in such a way that it minimizes the energy functional \cite{Xu}

198: 

199: \begin{equation}

200: E_s = \int_{0}^{1} (\frac{1}{2} \{\alpha {|{\bf C'}(s)|}^2 +

201: \beta {|{\bf C''}(s)|}^2 \} + E_{\rm ext} [{\bf C}(s)])\:\:ds

202: \end{equation}

203: 

204: \noindent

205: where the first two terms give the internal energy of the snake 

206: ($\alpha,\:\:\beta \geq 0$) and $E_{\rm ext}$ represents the external 

207: energy added to the system. ${\bf C'}$ and ${\bf C''}$ are the first

208: and second derivatives of the snake with respect to $s$. 

209: An example of external force for the snake is the gradient force:

210: $E_{\rm{ext}} = -{|{\bf \nabla} I(x,y)|}^2 $, where $ I(x,y) $ is the

211: image pixel intensity. In general, 

212: in the presence of an external force ${\bf V}({\bf C}(s,t))$, the time 

213: dynamics of such a snake is given by \cite{Ray}

214: 

215: 

216: \begin{equation}

217: \frac{\partial {\bf C}(s,t)}{\partial t} = \alpha {\bf C''}(s,t) - 

218: \beta {\bf C''''}(s,t) + {\bf V}({\bf C}(s,t)). 

219: \end{equation}

220: 

221: \noindent

222: Note that the snake in eqn. (2) is now a function of time $t$ as 

223: well. The stationary solution of eqn. (2) corresponds to a snake that 

224: minimizes the energy functional. Ray {\em et al} proposed an external force

225: $ V(x,y) $ obtained by solving the following partial differential 

226: equation (PDE) applying Dirichlet boundary condition[1]:

227: 

228: \begin{equation}

229: g(|{\bf \nabla}f|) {\nabla}^2 {\bf V} - (1 - g(|{\bf \nabla}f|)) ({\bf V} - 

230: {\bf \nabla}f) = {\bf 0},

231: \end{equation} 

232: 

233: \noindent

234: where $g(\alpha) = \exp(-K \alpha)$ and $f(x,y) = -E_{\rm ext}(x,y)$, 

235: $K$ being a tunable

236: parameter controlling the smoothness of the external snake force field.

237: Here we study the resultant dynamics due to 

238: the evolution of a number of such snakes, defined by the above

239: system of forces. 

240: 

241: When multiple snakes are evolved inside the desired closed boundary, 

242: the growth algorithm confirms that they all finally 

243: merge into a single snake after a finite time. This {\em merging of

244: snakes} is basically a topological effect and naively the

245: phenomenology reminds one of "nucleation" as seen in classical first-order

246: thermodynamic systems \cite{nucleation}. In our case, we allow a finite

247: number of GGVF snakes, each with a finite starting radius, to evolve in a 2-D 

248: plane and then numerically evaluate a few measurables -- the 

249: {\em nucleation time (NT)}, the {\em bounded area (BA)} after nucleation has

250: occurred, the 

251: {\em critical radius (CR)} at the time of nucleation and also the

252: {\em nucleation rate (NR)}, all as functions of snake evolution time. The 

253: respective quantities are defined as follows -- 

254: {\em nucleation time} is the time required for all the snakes to merge

255: together as a single unit,

256: {\em bounded area} is the sum of the areas of the initial snakes {\em before}

257: complete merging and the area under the single snake {\em after} merging, 

258: {\em critical radius} is the radius of curvature of the merged

259: structure and

260: {\em nucleation rate} is the ratio of the number of snakes to the

261: bounding area, before the merging has actually taken 

262: place. One should note that by {\em complete merging} we refer to the critical

263: phase when all the initial snakes merge together for the first time.

264: 

265: We perform two numerical experiments with merging. In the first

266: experiment, we start with a circular binary image of

267: radius $R$ containing $N$ number of circular snakes, each of radius

268: $r<R$, randomly distributed inside. 

269: As described before, the initial snakes are driven by GGVF

270: forces and they maintain a non-overlapping dynamics. In another

271: numerical experiment, we vary the radii of the smaller circles and later also

272: vary their total number. Both numerics are repeated with

273: varying sizes of the initial domain $R$. The enclosed figures 3 and 4 

274: show the variation of the bounding area and the critical radius

275: against time respectively, in non-dimensionalized units, in loglog plots. 

276: 

277: \begin{figure}

278: %\includegraphics[height=6.0cm,width=6.0cm,angle=0]{/home/akc/nilanjan/paper/fig3.ps}

279: \includegraphics[height=6.0cm,width=6.0cm,angle=0]{fig3.ps}

280: \caption{Bounding area vs time in log-log scale.} 

281: \label{Fig:3}

282: \end{figure}

283: 

284: \begin{figure}

285: \includegraphics[height=6.0cm,width=6.0cm,angle=0]

286: %{/home/akc/nilanjan/paper/fig4.ps}

287: {fig4.ps}

288: \caption{Critical radius vs time (log-log), after complete merging of the 

289: snakes.}

290: \label{Fig:4}

291: \end{figure}

292: 

293: \noindent

294: We find that all the graphs show excellent scaling in a broad domain, with 

295: the nonlinear zones referring to the saturation limits in each case. In the

296: figures shown, the bounding image radius is 80 units while the 

297: corresponding starting snake radius is 5 units. We have simulated using 

298: different number of initial snakes, indicated in the legend of the figures. 

299: The plots show power law

300: variations of each of the measured quantities with time, {\em i. e.} $y

301: \sim x^{\alpha}$ and the two relevant exponents, that of {\em critical radius} 

302: and {\em bounding area}, have the values  $\alpha_{\rm BA}

303: \approx 0.9-1.0,\:\:\alpha_{\rm CR} = 0.25$. We have used multiple

304: combinations of the parameter values but the exponents remain {\em universal},

305: that is they invariably follow a power law statistics. 

306: 

307: To analyse the results further, we begin with the

308: hypothesis that the phenomenology of the whole process, that of various snakes

309: {\em merging} together under certain boundary conditions in the

310: presence of suitably defined external forces, is fundamentally similar

311: to that of a domain growth process as encountered in the realms of

312: classical nucleation. In what follows, we try to evaluate the time rate of

313: evolution of a system of snakes before and after all

314: the snakes have completely merged with each other. We 

315: start from a model of solutes trying to nucleate in a solution and then

316: compare the results with the merging mechanism of snakes we find in the

317: AC method.

318: 

319: Our starting description is that of the diffusive growth of two

320: dimensional spherical droplets (circles, in 2-D) in a

321: Lifshitz-Slyozov \cite{nucleation, LStheory, binderstauffer} type of

322: continuum theory. We start with the thermodynamic definition of the work 

323: $ \Delta W $ required to form a nucleus from an aggregate of solute particles 

324: 

325: \begin{equation}

326: \Delta W = \delta (E - T_0 S + P_0 V)

327: \end{equation}

328: 

329: \noindent

330: where $ \delta (E - T_0 S) $ is the increment in the free energy of the

331: system and $ \delta (P_0 V) $ is the associated external work done.

332: Optimizing this total work factor, one can show 

333: \cite{nucleation} that the critical radius of a nucleating system, this being 

334: the rate of merger of snakes in AC method, is given by

335: 

336: \begin{equation}

337: R_{\rm CR} = \frac{\beta s'}{\mu'(P) - \mu_0'(P)}

338: \end{equation}

339: 

340: \noindent

341: Here $\mu_0'$ and $s'$ are the chemical potential and molecular

342: surface area of the nucleus, $\beta$ is the surface tension and $\mu'$

343: is the chemical potential of the solute in the solution. As more and

344: more solutes start nucleating {\em i. e.} snakes merge, the solution

345: approaches the critical limit of supersaturation and in the steady

346: state, the {\em critical radius} grows as

347: 

348: \begin{equation}

349: \frac{dR(t)}{dt} = D (\frac{\partial c}{\partial r})|_{r=R}

350: \end{equation}

351: 

352: \noindent

353: where $c(R)$ represents the spherically symmetric concentration distribution 

354: of snakes around a nucleus of radius $a$, $D$ being the diffusion 

355: coefficient.  

356: Following the treatment of Lifshitz-Slyozov \cite{nucleation,LStheory}, one 

357: can show that if $R_{\rm CR}(0)$

358: is the critical radius at the beginning of the merging process, then for

359: a diffusive nature of relaxation, the radius follows a dynamics 

360: 

361: \begin{equation}

362: \frac{dR(t)}{dt} = \frac{{R_{\rm CR}^2(0)}}{R(t)} (\frac{1}{R_{\rm CR}(t)} - 

363: \frac{1}{R(t)})

364: \end{equation}

365: 

366: \noindent 

367: We now define the dimensionless quantities $x(t) = \frac{R_{\rm CR}(t)}

368: {R_{\rm CR}(0)}$, $u(t) = \frac{R(t)}{R_{\rm CR}(t)}$ and 

369: $\tau = 2 \log x(t)$. The last one increases

370: monotonically from 0 to $\infty$ as the time $t$ increases likewise. 

371: Combining (6) and (7) in 2D, the dynamics is given by  

372: 

373: \begin{equation} 

374: \frac{d u^2(\tau)}{d{\tau}} = \gamma (1 - \frac{1}{u}) - u^2

375: \end{equation}

376: 

377: \noindent

378: where $\gamma = \frac{dt}{R_{\rm CR}(0) dx} > 0$. A linear stability analysis 

379: of the above nonlinear equation gives a fixed point at $\gamma_0 = 

380: \frac{27}{4}$. We are interested in the dynamics around an

381: $\epsilon$-neighborhood of this point $\gamma_0$. If $\gamma(\tau)=

382: \frac{27}{4}[1 - {\epsilon}^2(\tau)]$ ($\epsilon \rightarrow 0$ as 

383: $\gamma \rightarrow \infty$), then near the critical point 

384: ($ u_0 = {(\gamma/2)}^{1/3} = 3/2 $) the merging snakes follow a 

385: dynamics defined by   

386: 

387: \begin{equation}

388: \frac{du}{d{\tau}} = -2 {(u - \frac{3}{2})}^2 - \frac{3}{4} {\epsilon}^2

389: \end{equation}

390: 

391: \noindent

392: The time variation of the merging nuclei (snakes) is

393: 

394: \begin{equation}

395: x(t) = \frac{4}{27} \frac{\sqrt{t}}{R_{\rm CR}(0)}

396: \end{equation}

397: 

398: \noindent

399: Thus the bounding area ($\sim x^2(t)$) of the merging snakes

400: grow at the rate of $t$ which is roughly speaking our numerical estimate 

401: (0.9-1.0) also

402: (Fig. 3). However, the above analysis is only valid {\em before} complete

403: merging of the snakes has occurred. In the {\em merged} phase when 

404: the resultant  

405: asymmetrical structure continues growing finally to coalesce with the bounding

406: image, the system dynamics is modified. To analyse the situation,

407: we start from eqn.(7). The equation may alternatively be represented as

408: 

409: \begin{equation}    

410: \frac{dR(t)}{dt} = \frac{D \delta c(R(t))}{R(t)} \sim \frac{1}{T} \

411: \frac{1}{R^2(t)}.

412: \end{equation}

413: 

414: \noindent

415: $T$ in the above equation is the temperature of the solution in the

416: equivalent thermodynamic problem which in our case, is a measure of the 

417: average surface energy $E$, $ E \propto T $, of the system. 

418: Evidently, in 2D, $E$ goes as $2 \pi R \beta$. 

419: Plugging these values in the above

420: equation, we see that after complete merging of the snakes has taken place,

421: the effective radius of curvature of the resultant structure grows as 

422: $R_{\rm CR}(t) \sim t^{1/4}$ (Fig. 4). Once again our numerical result 

423: is in exact harmony with the theory.

424: 

425: Our above analysis, both numerical and analytical clearly suggests that below 

426: the apparently simplistic level of the GGVF application, the system dynamics 

427: has a more fundamental symmetry. This symmetry comes from  

428: the fact that the GGVF method lies in the same

429: universality class as that of a classical nucleation model. This first-order

430: critical response of the system is what 

431: provides the subtlety of the underlying physics. There is, however,

432: a notable shortcoming with the GGVF technique, that it does 

433: not allow us the liberty of starting with an initial condition at an 

434: arbitrary location. If the

435: snake starts at a position in which a major portion of the initial snake is

436: outside the desired boundary or {\em vice versa}, 

437: then the snake driven by GGVF will not converge to the

438: actual boundary. Although it can be shown by the Reed-Simon's

439: theorem \cite{reedsimon} that a convex set (a circle, say) growing within 

440: a larger convex set (a larger circle or rectangle, say) will always merge 

441: with the outer boundary under the action of isotropic driving forces, to 

442: the best of our knowledge, no such mathematical lemma exists for a

443: convex set growing in a concave set or vice versa. 

444: 

445: In summary, our achievement in the present paper

446: has been to analyze the physical foundation of the GGVF merging technique

447: and to provide answer to the rather puzzling question as to why it

448: works so accurately. In the process, we have shown that 

449: the answer lies in the general scaling behavior of the underlying 

450: nucleation dynamics, defined by proper scaling laws. This clearly 

451: indicates that an active contour system is in the same universality class as the

452: nucleation model we considered. Our results, in unison with probable

453: biomedical applications, are expected to inspire further 

454: studies in the understanding of lung-based diseases.

455: 

456: 

457: \begin{thebibliography}{99}

458: 

459: 

460: 

461: \bibitem{Ray} N. Ray, S. T. Acton, T. Altes, E. E. de Lange and 

462: J. R. Brookeman, IEEE Trans. Med. Imaging {\bf 22}, 189 (2003). 

463: 

464: \bibitem{Xu} C. Xu and J. L. Prince, Signal Processing {\bf 71}, 

465: 131 (1998).

466: %\bibitem{Tally} Tally

467: 

468: \bibitem{Kass} M. Kass, A. Witkin and D. Terzopolous, Int. J. Comput. Vis.

469: {\bf 1}, 321 (1987).

470: 

471: %\bibitem{Ray1} N. Ray, {\em et al}  

472: 

473: %\bibitem{detector} Detector

474: 

475: \bibitem{Canny} J. F. Canny, IEEE Trans. Pattern Analysis and Machine 

476: Intelligence, 679 (1986).

477: 

478: \bibitem{McInerney} T. McInerney and D. Terzopolous, 

479: Med. Image Anal. {\bf 4}, 73 (2000).

480: 

481: %\bibitem{Ray02} N. Ray, {\em et al}  

482: 

483: %\bibitem{Ray03} N. Ray, {\em et al} IEEE Trans. Med. Imaging {\bf 22},

484: 

485: %page (2003).

486: 

487: \bibitem{nucleation} J. D. Gunton, M. San Miguel and P. S. Sahni in

488: {\em Phase transitions and critical phenomena}, edtd. by C. Domb and

489: J. L. Lebowitz, vol. 8, Academic Press (1983); E. M. Lifshitz and

490: L. P. Pitaevskii in {\em Physical Kinetics}, Landau and Lifshitz

491: Course of Theoretical Physics, vol. 10, Butterworth-Heinemann, 1981.

492: 

493: \bibitem{LStheory} I. M. Lifshitz and V. V. Sloyozov, J. Phys. Chem

494:   Solids {\bf 19}, 35 (1961); C. Wagner, Z. Electrochem. {\bf 65}, 581

495: (1961).

496: 

497: \bibitem{binderstauffer} K. Binder and D. Stauffer, Adv. Phys. 

498: {\bf 25}, 343 (1976).

499: 

500: \bibitem{Langer69} J. S. Langer, Ann. Phys. {\bf 41}, 108 (1967);

501: Ann. Phys. {\bf 54}, 258 (1969).

502: 

503: \bibitem{Langer73} J. S. Langer and L. A. Turski, Phys. Rev. A {\bf 8}, 

504: 3230 (1973); {\em ibid}, Phys. Rev. A {\bf 22}, 2189 (1980).

505: 

506: \bibitem{reedsimon} M. Reed, B. Simon and S. M. Reed in the "Methods of

507: Modern Mathematical Physics: Fourier Analysis, Self-adjointness", Academic

508: Press, 1997.

509: 

510: \end{thebibliography}

511: 

512: \end{document}

513: 

514: