Robust Computation of Mutual Information Using Spatially Adaptive Meshes 1st Edition by Hari Sundar, Dinggang Shen, George Biros, Chenyang Xu, Christos Davatzikos ISBN 9783540757573 - Own the ebook now with all fully detailed chapters
Robust Computation of Mutual Information Using Spatially Adaptive Meshes 1st Edition by Hari Sundar, Dinggang Shen, George Biros, Chenyang Xu, Christos Davatzikos ISBN 9783540757573 - Own the ebook now with all fully detailed chapters
_____ Follow the link below to get your download now _____
https://ebookball.com/product/robust-computation-of-mutual-
information-using-spatially-adaptive-meshes-1st-edition-by-
hari-sundar-dinggang-shen-george-biros-chenyang-xu-christos-
davatzikos-isbn-9783540757573-9534/
https://ebookball.com/product/elements-of-the-theory-of-
computation-2nd-edition-by-harry-lewis-christos-
papadimitriou-0132624788-9780132624787-10788/
Modeling of Needle Tissue Interaction Using Ultrasound
Based Motion Estimation 1st Edition by Ehsan Dehghan, Xu
Wen, Reza Zahiri Azar, Maud Marchal, Septimiu E Salcudean
ISBN 9783540757573
https://ebookball.com/product/modeling-of-needle-tissue-interaction-
using-ultrasound-based-motion-estimation-1st-edition-by-ehsan-dehghan-
xu-wen-reza-zahiri-azar-maud-marchal-septimiu-e-salcudean-
isbn-9783540757573-10816/
Abstract. We present a new method for the fast and robust computation of in-
formation theoretic similarity measures for alignment of multi-modality medical
images. The proposed method defines a non-uniform, adaptive sampling scheme
for estimating the entropies of the images, which is less vulnerable to local max-
ima as compared to uniform and random sampling. The sampling is defined using
an octree partition of the template image, and is preferable over other proposed
methods of non-uniform sampling since it respects the underlying data distribu-
tion. It also extends naturally to a multi-resolution registration approach, which
is commonly employed in the alignment of medical images. The effectiveness of
the proposed method is demonstrated using both simulated MR images obtained
from the BrainWeb database and clinical CT and SPECT images.
1 Introduction
Inter-modality image alignment is a fundamental step in medical image analysis. It is
required to bring image data from different modalities to a common coordinate frame
in order to accumulate information. It is usually presented as an optimization prob-
lem requiring the minimization of a certain objective function. Objective functions, or
similarity measures based on information theoretic principles have been successful in
aligning images from differing modalities. Mutual Information (MI) was proposed as
an image similarity measure by Collignon [1], Viola [2] and Wells [3] and is widely
used for rigid inter-modality registration. Several modifications have been proposed to
make MI more robust and increase its capture range, including Normalized Mutual In-
formation [4]. However, MI-based methods are very sensitive to the way the probability
distributions are estimated and the accuracy of the estimated probability distributions
have a great influence in the accuracy and robustness of the registration results [5].
A common assumption made in estimating the probability distribution is that each
voxel is an i.i.d. realization of a random variable. Therefore, the probability distributions
(including the joint distribution) can be computed by using all voxels in the reference
image and the corresponding voxels in the transformed subject image. In general this
can be quite expensive to compute and several multi-resolution and sampling techniques
have been proposed for faster estimation of the probability distribution. Downsampling,
both uniform and random, have been used quite commonly to speed up the estimation
of the distributions [5]. Nonlinear sampling techniques where the local sampling rate is
proportional to the gradient magnitude have also been proposed [6].
N. Ayache, S. Ourselin, A. Maeder (Eds.): MICCAI 2007, Part I, LNCS 4791, pp. 950–958, 2007.
© Springer-Verlag Berlin Heidelberg 2007
Robust Computation of Mutual Information Using Spatially Adaptive Meshes 951
In general, although these methods have performed quite well for registering differ-
ent structural modalities (like CT and MR), they have been less successful in being able
to register structural modalities to functional modalities,which is important for diagno-
sis and treatment planning applications. Functional modalities do not image all tissues
and are therefore more sensitive to errors in the estimation of probability distributions.
We shall use the example of registering Single Photon Emission Computed Tomogra-
phy (SPECT) with CT images of the same patient to demonstrate this problem.
In this paper we present a new method for rigid alignment of multi-modality images
using an octree based partitioning of the reference image. Octrees allow us to parti-
tion the image into spatially adaptive regions (octants) such that homogeneous regions
produce large octants. The octree allows us to define a non-linear sampling of the im-
age that is proportional to the underlying image complexity. The samples thus obtained
are closer to the usual i.i.d. assumption in that they tend to be less statistically inter-
dependent, which in turn help us obtain more accurate estimates of entropy. The MI is
the sum of the entropy of the subject image and the entropy of subject conditional on
the target. Consequently, improved entropy estimates provide better estimates of MI.
The rest of the paper is organized as follows. In Section 2 we present a brief introduc-
tion to estimating entropy in images, and lay the foundation for our arguments in favor
of the octree-based estimation of distributions, which is described in Section 3. Section
4 discusses the incorporation of the octree-based mutual information metric into a reg-
istration framework for inter-modality alignment of images, within a multi-resolution
framework. Experimental results and comparisons with other methods are provided in
Section 5.
Δ
n
H(X) = − pi log pi . (1)
i=0
Fig. 1. The difficulty in estimating the entropy of images. (a) An image with the corresponding
histogram and entropy estimate, (b) An image with objects moved within the image to make it
appear more random than (a). The standard estimate for the entropy does not change, since the
histogram is still the same. (c) Octree-based entropy estimation for the image shown in (a), and
(d) octree-based entropy estimation for the image shown in (b). Note that in this case the increase
in entropy was captured by the octree.
objects in the scene, the entropy is unaffected. This happens because of our assumption
that each voxel intensity is an independent sample. This is not true since voxel inten-
sities depend on the underlying objects, and samples from the same object cannot be
assumed to be independent. Also observe that the gradient based approaches will not be
able to capture this difference either, because it is not affected by the spatial variation
between the two configurations shown in Figures 1(a) and (b).
It is widely accepted that images can be successfully modeled as Markov random
fields [8,9]. Now, assuming that instead of an i.i.d. assumption, the samples form a
Markov random field, then the error in the estimation of the entropy is lowered if the
samples are largely independent of each other. Sabuncu et al. [6] suggest two non-
linear sampling strategies based on the magnitude of the image gradient, in order to
make the samples more independent. These suggested sampling methods, however, are
for 2D images, expensive to compute and not easily extensible for 3D images. Algo-
rithms have been proposed to make mutual information more robust by incorporating
geometric information using gradients [10] and image salience [11]. However, the gra-
dient captures only the local image complexity and it is not straightforward to extend
this within a scale independent framework. This suggests using an adaptive sampling
strategy, and adaptive image representations have been commonly used in non-linear
image registration for representing the transformation [12] and for estimating local im-
age similarities. However, to the best of our knowledge, it has not been used to estimate
the global similarity of images.
If we can define a partitioning on the image that is spatially adaptive, and creates largely
independent regions, we will be able to use the partition to define a better sampling
strategy. Binary space partitions (BSP) [13] allow us to generate such spatially adaptive
Robust Computation of Mutual Information Using Spatially Adaptive Meshes 953
Fig. 2. (a) An example of a quadtree generated from a 2D slice of a MR brain image demonstrat-
ing how the quadtree adapts to the underlying data, (b) an example of a quadtree generated from
a 2D slice of a thoracic CT image, and (c) the quadtree generated from a cardiac short axis MR
image
where bin(·) defines the histogram binning operator and Oct(T ) is the octree computed
for the template image T . In order to understand why an octree-based sampling appears
to be better than uniform or gradient based sampling, consider the example discussed
954 H. Sundar et al.
earlier, shown in Figure 1. We showed earlier that traditional sampling methods (both
uniform and gradient based) estimate the same entropy for the images shown in Figures
1(a) and (b). However, the octree-based sampling is able to capture this difference, as
seen in Figures 1(c) and (d). This is because the spatial complexity of the octree in-
creases as a result of the change in the randomness of the scene. The octree captures
the spatial complexity of the scene and consequently the entropy of a complex scene is
higher as it has a denser octree. It is important to point out that the octree is not necessar-
ily the best estimate of spatial complexity, since it is partial towards objects that are axis
aligned and will not be able to capture such variations. A better estimate of the scene
complexity would be a connectivity graph of segmented objects. This however would
increase the complexity of both computing such a graph and also in using it during
the registration. It would be difficult and computationally expensive to define transfor-
mations and interpolation functions on a connectivity graph of segmented tissues. The
octree is a compromise that provides a sufficiently accurate and robust estimate of the
entropy (see Section 5) and is also easy to compute and manipulate. Interpolation and
transformations (linear and non-linear) can be easily defined on the octree. Importantly,
octree-representations are amenable to parallel computing, which can dramatically ex-
pedite the performance of algorithms that use it. In all the experiments reported in this
paper, the octree was computed only for the template image and each octant was sam-
pled at the center of each octant in the template image. Although, it would be better to
average over the entire octant instead of sampling at the center, we opted for the latter
to improve the computational efficiency of the method.
Ê
Given a template image, T : Ω → n and a subject image S : Ω → n , where Ê
Ê
Ω ∈ d , the goal is to determine a rigid transformation χ that aligns the two images.
The similarity measure is the metric that quantifies the degree of alignment and allows
us to present the alignment problem as an optimization problem. We use the octree-
based similarity measures as described in Section 3. We formulate the determination of
the optimal transformation, χ̂, as an optimization problem:
where, I(.; .) is the Mutual Information. Powell’s multidimensional set method [20] is
used to iteratively search for the maxima along each parameter using Brent’s method
[20]. In order to increase the robustness and to improve speed, we use a multi-resolution
framework defined on the octree. The octree at lower (coarser) resolutions are generated
by skipping all octants at the finer levels.
5 Results
In this section we describe experiments that were carried out to test the effectiveness
of octree-based MI in the rigid registration of inter-modality images. We first describe
Robust Computation of Mutual Information Using Spatially Adaptive Meshes 955
Fig. 3. Comparison of the mutual information computed via uniform sampling (dotted lines) and
using the proposed octree-based sampling (solid lines), on BrainWeb datasets. The plots shown
are for a comparison between a T1-weighted (T1) image and a proton density (PD) image with
9% noise and 40% intensity non-uniformity.
the similarity profiles when an artificial transformation is introduced between two reg-
istered images. We compared the octree-based method with uniform sampling based
estimation of mutual information. The first experiment was performed using simulated
MR datasets obtained from the BrainWeb database [21]. The second experiment was
performed with 13 CT datasets with corresponding SPECT images. These images were
all acquired using a Siemens Symbia™ T, a TruePoint SPECT-CT system and are as-
sumed self registered. We analyzed the mutual information profiles while varying the
transformation. The transformation parameters were varied one at a time, and the sim-
ilarity profiles were plotted. The plots for translation along the x-axis, and for rotation
about the x and y axes are shown in Figures 3 and 4, for T1-PD MR images and CT-
SPECT images, respectively1. The profiles for translation and rotation along the other
axes were similar. In all cases we compare the octree-based sampling with uniform sam-
pling, where the total number of samples are similar. The octree reduced the number of
samples by a factor of 8 on an average, therefore we subsampled by a factor of 2, along
each direction, for the uniform sampling strategy, to have the same number of samples
in both cases. As can be seen from Figure 3, both methods perform equally well on the
BrainWeb datasets. Both sampling techniques have smooth curves with sharp peaks and
very good capture ranges. However, when we look at the results from the CT-SPECT
comparison, shown in Figure 4, we observe that the octree-based sampling performs
much better. Although, both approaches have good profiles subject to translation, for
the profiles subject to rotation, the uniform sampling approach exhibits a weak max-
ima at the optimal value with a very small capture range. In contrast the octree-based
approach exhibits a strong maximum at the optimal value and also has a much larger
capture range. The fact that the neighboring maxima in the vicinity of the optimum are
lower further implies that it is likely that a multi-resolution approach can potentially be
used to increase the capture range. The uniform sampling approach will in most cases
converge to the wrong result since the neighboring maxima are much larger in that case.
1
The profiles generated from using all the voxels in the image were almost identical to those
obtained by uniform subsampling, and are not presented here for clarity.
956 H. Sundar et al.
Fig. 4. Comparison of the mutual information computed via uniform sampling (dotted lines)
and using the proposed octree-based sampling (solid lines), with CT-SPECT datasets. The plots
shown are for a comparison between a CT cardiac image (512 × 512 × 25) and a SPECT image
(128 × 128 × 128).
Table 1. Means and standard deviations of the registration errors for the different test cases
6 Conclusion
We have presented a spatially adaptive sampling method for the estimation of image
entropies and mutual information. We also demonstrate the improvements in the rigid
registration of CT and SPECT images using the new sampling method. The proposed
sampling offers flexibility between robustness and speed. When used at full resolution,
octree-based sampling provides better estimates of image entropy and mutual infor-
mation between images. In addition, it is better to use octree-based sampling in order
to speed up the estimation of the similarity measure as opposed to using uniform sam-
pling approaches. When compared with uniform sampling approaches, the octree-based
sampling was more accurate for comparable computational speeds, and appears to be a
better method for speeding up the estimation of MI.
Acknowledgments
The authors would like to thank Dr. Parmeshwar Khurd for useful discussions. This
research was partially funded by a research grant from Siemens Corporate Research,
Princeton, NJ.
References
1. Collignon, A., et al.: Automated multimodality medical image registration using information
theory. Proc. Information Processing in Medical Imaging 3, 263–274 (1995)
2. Viola, P.: Alignment by Maximization of Mutual Information. PhD thesis, Massachusetts
Institute of Technology (MIT), Cambridge, Massachusetts (1995)
3. Wells, W., Viola, P., Atsumi, H., Nakajima, S., Kikinis, R.: Multi-modal volume registration
by maximization of mutual information. Medical Image Analysis 1(1), 35–51 (1996)
4. Studholme, C., Hill, D., Hawkes, D.: An overlap invariant entropy measure of 3D medical
image alignment. Pattern Recognition 32(1), 71–86 (1999)
5. Pluim, J., Maintz, J., Viergever, M.: Mutual-information-based registration of medical im-
ages: a survey. IEEE Trans. on Medical Imaging 22(8), 986–1004 (2003)
6. Sabuncu, M.R., Ramadge, P.J.: Gradient based nonuniform sampling for information theo-
retic alignment methods. In: Proc. Int. Conf. of the IEEE EMBS, San Francisco, CA, IEEE
Computer Society Press, Los Alamitos (2004)
7. Shannon, C.E.: A mathematical theory of communication. The Bell System Technical Jour-
nal 27 (1948)
8. Li, S.Z.: Markov random field modeling in computer vision. Springer, Heidelberg (1995)
9. Prez, P.: Markov random fields and images. CWI Quarterly 11(4), 413–437 (1998)
10. Pluim, J.P.W., Maintz, J.B.A., Viergever, M.A.: Image registration by maximization of com-
bined mutual information and gradient information. IEEE Trans. on Medical Imaging 19(8),
809–814 (2000)
11. Luan, H., Qi, F., Shen, D.: Multi-modal image registration by quantitative-qualitative mea-
sure of mutual information Q-MI. In: Proc. CVBIA, pp. 378–387 (2005)
12. Timoner, S.: Compact Representations for Fast Nonrigid Registration of Medical Images.
PhD thesis, Massachusetts Institute of Technology (MIT), Cambridge, Massachusetts (2003)
13. Fuchs, H., Kedem, Z.M., Naylor, B.F.: On visible surface generation by a priori tree struc-
tures. In: Proc. SIGGRAPH, pp. 124–133 (1980)
958 H. Sundar et al.
14. Finkel, R.A., Bentley, J.L.: Quad trees: A data structure for retrieval on composite keys. Acta.
Inf. 4, 1–9 (1974)
15. Meagher, D.: Geometric modeling using octree encoding. Computer Graphics and Image
Processing 19, 129–147 (1982)
16. Laferté, J., Perez, P., Heitz, F.: Discrete Markov image modeling and inference on the
quadtree. IEEE Trans. on Image Processing 9(3), 390–404 (2000)
17. Bouman, C.A., Shapiro, M.: A multiscale random field model for Bayesian image segmen-
tation. IEEE Trans. on Image Processing 3(2), 162–177 (1994)
18. Luettgen, M.R., Karl, W.C., Willsky, A.S.: Efficient multiscale regularization with applica-
tions to the computation of optical flow. IEEE Trans. on Image Processing 3(1), 41–64 (1994)
19. Irving, W.W., et al.: An overlapping tree approach to multiscale stochastic modeling and
estimation. IEEE Trans. on Image Processing 6(11), 1517–1529 (1997)
20. Press, W.H., Teukolsky, S.A., Vetterling, W.T., Flannery, B.P.: Numerical Recipes in C: The
Art of Scientific Computing. Cambridge University Press, New York (1992)
21. Collins, D.L., et al.: Design and construction of a realistic digital brain phantom. IEEE Trans.
on Medical Imaging 17(3), 463–468 (1998)
Other documents randomly have
different content
The Project Gutenberg eBook of Il fantasma di
Canterville e il delitto di Lord Savile
This ebook is for the use of anyone anywhere in the United States
and most other parts of the world at no cost and with almost no
restrictions whatsoever. You may copy it, give it away or re-use it
under the terms of the Project Gutenberg License included with this
ebook or online at www.gutenberg.org. If you are not located in the
United States, you will have to check the laws of the country where
you are located before using this eBook.
Illustrator: G. Mazzoni
Language: Italian
Il fantasma di Canterville
e il delitto di Lord Savile
Prima versione italiana di G. Vannicola.
con disegni di G. Mazzoni.
Seconda Edizione.
*
* *
«Ciò che il paradosso era per me nella sfera del pensiero — dice
Wilde nel De profundis — la perversità lo divenne nel dominio della
passione».
Il «paradosso» non è altro, insomma, che una verità poco familiare e
che il tempo attenuerà in verità usuale e, forse, in luogo comune: il
nome che gl'imbecilli danno alla verità — diceva Jean Moréas,
quando lo accusavano d'esser paradossale.
Alcune «verità poco familiari» sono una fra le più notorie
caratteristiche dell'opera di Oscar Wilde. Frasi nette, lucide, boutades
lanciate col piccolo colpo secco di una tabacchiera che si richiude:
— Nessun delitto è volgare. Ma ogni volgarità è delitto. La volgarità è
la condotta degli altri.
— Si dovrebbe esser sempre un poco inverosimili.
— Esser prematuro, significa esser perfetto.
— Una verità cessa di esser vera quando più di uno crede in lei.
— Soltanto gli dei conoscono la morte. Apollo è scomparso. Ma
Giacinto il quale, secondo gli uomini, venne sgozzato da lui, vive
ancora: Nerone e Narciso son sempre con noi.
— La condizione della perfezione è la pigrizia. Lo scopo della
perfezione è la giovinezza.
— Evitate gli argomenti di non importa qual genere. Essi sono
sempre volgari e spesso convincenti.
E questa definizione delle donne:
— Sfingi senza segreto.
E questo aforisma in difesa dell'egoismo:
— Il mezzo sicuro di non conoscer nulla della vita, è quello di cercare
d'essere utile.
Wilde amava suscitare il riso, sorridendo; ma si compiaceva anche
ad una specie di emozione quasi ostile al riso, la cui qualità potrebbe
definirsi «opulenza», magnificenza, magistero di arte che ordisce la
trama con fila d'oro e la ricama con gemme.
Se non precisamente un classico del ridere, Wilde è un classico di
quell'humour così particolare agl'inglesi, cui egli aggiunge un sapore
di decadenza singolarmente acconcio all'anima pagana che l'invade e
lo tormenta:
«Quando Gesù volle rientrare in Nazaret, egli narrava, Nazaret era
così cambiata che Gesù non riconobbe più la sua città. La Nazaret
ove egli aveva vissuto era piena di lamentazioni e di lagrime, questa
città era piena di risa e di canti. E Cristo, entrando in città, vide degli
schiavi carichi di fiori affrettarsi verso la scalea di una casa di marmo
bianco. Cristo entrò nella casa, e in fondo ad una sala di diaspro,
coricato sopra un giaciglio, vide un uomo i cui capelli disfatti erano
mischiati alle rose rosse e le cui labbra erano rosse di vino.
Cristo si avvicinò a lui, gli toccò la spalla e gli disse: — Perchè
conduci questa vita? — L'uomo si volse, lo riconobbe e rispose: —
Ero lebbroso; tu m'hai guarito. Perchè condurrei un'altra vita?
Cristo uscì da quella casa. Ed ecco che nella strada vide una donna il
cui viso e le vesti erano dipinti, e i cui piedi erano calzati di perle; e
dietro di lei camminava un uomo il cui abito era di due colori e i cui
occhi si gravavano di desiderio. E Cristo si avvicinò all'uomo, gli toccò
la spalla e gli disse: — Perchè dunque segui questa donna e la
guardi così? — L'uomo si volse, lo riconobbe e rispose: — Ero cieco;
tu m'hai guarito. Che altro farei della mia vista?
E Cristo si avvicinò alla donna: — La strada che tu segui, le disse, è
quella del peccato; perchè seguirla? — La donna lo riconobbe e gli
disse ridendo: — La strada ch'io seguo è gradevole, e tu hai
perdonato tutti i miei peccati.
Allora Cristo sentì il suo cuore colmo di tristezza e volle abbandonare
questa città. Ma come ne usciva, vide infine, seduto sull'orlo dei
fossati della città, un giovine che piangeva. Cristo gli si appressò e
toccando le ciocche dei suoi capelli gli disse:
— Amico mio, perchè piangi?
Il giovine levò gli occhi, lo riconobbe e rispose: — Ero morto e tu
m'hai risuscitato; che altro farei della mia vita?»
Non è agevole cosa definire la qualità del riso di Wilde. È un ridere
leggero, un condurre di prato in prato relegante armento di delicate
«fumisteries», immaginate e dette su fumo di sigarette.
Di questo suo humour personalissimo diamo esempio, in questa
raccolta, con la traduzione dei due deliziosi etchings che seguono,
racconti di buffoneria, dove Wilde, come sempre, rimane serio.
G. Vannicola.
IL FANTASMA DI CANTERVILLE
I.
Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.
ebookball.com