UsimagTool: an Interactive Research Tool for Ultrasound Image Processing
Rub´en C´ardenes, Antonio Trist´an, Lucilio Cordero-Grande,
Emma Mu˜noz-Moreno and Marcos Mart´ın-Fern´andez
Laboratory of Image Processing, ETSI Telecomunicaci´on
University of Valladolid, Spain
{
ruben,atriveg,lcorgra,emunmor,marcma
}
@lpi.tel.uva.es
Abstract
We present here an open source software package called UsimagTool which consist of a set of algorithms and a graphic user interface (GUI) specially designed for re-search purposes in Ultrasound image processing. This tool is based on the C++ open source libraries ITK, VTK and FLTK. One of its main features is its modular and flexible architecture, which allow developers to modify or include image processing algorithms very fast.
UsimagTool includes several algorithms developed specifically for Ultrasound data processing, that has been coded using the ITK and VTK convention. The algorithms included and described here are new implementations and they are not part of the official ITK or VTK releases. They are three filters (anisotropic Wiener, Speckle Reduction Anisotropic Diffusion, and Tensor guided Anisotropic Diffu-sion) one segmentation algorithm based on a Markov ran-dom field model, and a demons registration suitable for US data.
There is also a user friendly GUI designed for segmenta-tion and registrasegmenta-tion tasks, with three independent viewers to display simultaneously three different data sets. The GUI can be easily adapted or modified with little effort for other tasks, for instance it can be redesigned for clinical use.
1
Introduction
The extraordinary growth experimented by the medical image processing field in the last years, has motivated the development of many algorithms and software packages for image processing. One of the most important set of algo-rithms for image processing can be found in the Insight Toolkit (ITK) [11] open source libraries for segmentation and registration of multidimensional data. At the same time, several software packages for medical image visualization and analysis have been developed using ITK and the visu-alization toolkit (VTK) [26] libraries. There are also some
efforts to develop software that could be easy to modify by other researchers, for example the Medical Imaging Interac-tion Toolkit (MITK) [32], is intended to fill the gap between algorithms and the final user, providing interaction capabil-ities to construct clinical applications.
Therefore, many software packages for visualization and analysis of medical data are available for the research munity. Among them we can find commercial and non com-mercial packages. Usually, the former ones are focused on specific applications or just on visualization being more ro-bust and stable, whereas the latter ones often offer more features to the end user, and they are usually available as open source projects. Some of the non commercial pack-ages equipped with graphic user interface (GUI), are 3D Slicer [21, 10], Julius [24], Osirix [25], ITK-SNAP [34], Medical Studio [30], MedInria [8], Amide [15] and Fusion-Viewer [16], to cite some of them. Other important com-mercial software packages are Analyze [19], Amira [20] and 3D Doctor [2].
The requirements that we look for in a medical imaging software package for researchers can be summarized in the following list:
• Open source code: to be able for everyone to modify and reuse the source code.
• Efficiency, robust and fast: using a standard object ori-ented language such as C++.
• Modularity and flexibility for developers: in order to change or add functionalities as fast as possible. • Multi-platform: able to run in many Operating systems
to be useful for more people.
• Usability: provided with an easy to use GUI to interact as easy as possible with the end user.
• Documentation: it is important to have a well docu-mented web site as well as a complete user and devel-oper manuals or tutorials.
2D and 3D Image Readers
Writers and Viewers 3D Viewers
ITK
FLTK
VTK
UsimagTool
Buttons . windows, menus
Image Processing 3D visualization
Figure 1. UsimagTool Architecture
Among the non commercial packages mentioned before, Amide and FusionViewer do not accomplish all these fea-tures because they are not specially designed for developers, and they are not well documented for that, Osirix is only available for Mac OS, MedInria is not open source, Med-ical Studio is in an early stage of development, and only ITK-SNAP, Slicer and Julius accomplish almost all of the features mentioned, but their structure is relatively complex compared with the architecture of UsimagTool, thus they do not offer a flexible architecture for researchers. We propose here a package that fulfills all these requirements and at the same time is equipped with a set of algorithms for US image processing, which makes this package one of the few soft-ware packages developed for Ultrasound image processing. The paper is structured as follows: next section de-scribes the architecture of UsimagTool, section 3 dede-scribes the graphic user interface and the visualization capabilities. Then we explain in section 4 the processing algorithms im-plemented, and finally we present our conclusions and fu-ture work.
2
Architecture
UsimagTool is built over the C++ libraries Insight Toolkit (ITK) (that implements image processing routines), the Visualization Toolkit (VTK) libraries (that implements 3D visualization) and the Fast Light Toolkit (FLTK) li-braries [27] (that implements graphic objects), as shown in figure 1. It has three main classes: UsimagToolBase, the base class that defines and declares the basic objects and types (filters, images and viewers), UsimagToolGUI which is inherited from the previous class and that implements all the graphic objects (windows, buttons, menus, etc.) and the class UsimagToolConsole, which is also inherited from the previous one and that implements the callbacks of the but-tons declared in the UsimagToolGUI class. These callbacks set up the parameters of the image processing functions, and connect and execute the pipeline between the input and the output. The scheme of classes is shown in figure 2.
This simple architecture allows a rapid and modular de-sign of new functionalities in our software package.
More-3D Viewers (VTK) (FLTK) 2D Viewers Image Processing Libraries (ITK) UsimagToolGUI UsimagToolBase
UsimagToolConsole Callbacks implementations Graphical objects and parameters definitions and declarations Basic objects, functions and types definitions and declarations
Figure 2. UsimagTool class diagram
Figure 4. Dialog windows to set up the param-eters of the Wiener filter (left) and the demons registration (right)
over, the graphic part is designed using an open source tool called fluid which is an easy and intuitive program to design GUIs in FLTK.
The algorithms for image processing are included as li-brary modules, most of them are already included in the ITK libraries, and they are compiled and linked with the main interface separately. The process to include a new al-gorithm to the package is relatively simple, once the ITK class code is done. Mainly three steps are needed, one for every class: first we need to include and declare the new type and object associated to our new algorithm in the base class. The second step is to include the graphic objects in the GUI class, to interact with the user to introduce the pa-rameters and execute the algorithms. Finally we include in the console class the callback code called from the previous class that connects the pipeline from the input to the output.
3
Graphic User Interface
A screen-shot of the GUI is shown in figure 3. The GUI is divided into three viewers, a control panel at the left side with input/output controls, and functionalities divided into tabs, a menu, a state bar that indicates the percentage of progress of algorithms being executed, and several quick ac-cess buttons (Open, 3D viewer, auxiliary result viewer and exit).
Figure 3. UsimagTool Graphic User Interface
We also show some dialog windows in figure 4 designed to interact with the user to set up the parameters of the se-lected function, in this figure an anisotropic Wiener filter and a demons registration.
3.1
Visualization
Visualization is performed in three main viewers as men-tioned above. Each viewer shows two dimensional slices from different 3D data, and support the following functions: zoom, display any of the three orthogonal views, flip the x,y, or z axis, transpose the axis of the slice being viewed, dis-play points selected, show image details, view a color over-lay image, show pixel value and location of cursor, change intensity window and level, and switch between different visualization modes.
Our experience tell us that for research purposes it is im-portant to see at the same time different data sets, for in-stance the original data, the filtered data and the segmented data in the case of a segmentation pipeline, or the source the target and the registered data in the case of a registration algorithm. For this reason we have three separate viewers to show three different data sets. For every process, the user decides the input image and the output viewer in the upper part of the control panel. The sliders of two viewers syn-chronizes automatically when the data have the same size in any two viewers, thus allowing a better comparison of the data. The user can also use an auxiliary viewer (in a sep-arate window) with the same visualization features as the main viewers, (see figure 5) and also a 3D viewer to show a volume rendering of the data.
4
UsimagTool Functionalities
UsimagTool provides input/output support for several
Figure 5. Auxiliary viewer
data formats: DICOM, MetaHeader, Raw Data, GE Volu-son, and 2D formats such as TIFF, PNG, and JPG. Notice that Voluson data format is used for US, which is not usually supported by other software packages.
Regarding the image processing functionalities of UsimagTool we have divided them into four categories: Ba-sic Operations, Filtering, Segmentation, and Registration, that are organized into tabs in the GUI control panel. We will explain now in detail the US image processing algo-rithms developed in our tool.
4.1
Basic Operations
In order to support pre and/or post processing of the data images, some basic operations are included in this tool. They include gradient magnitude computation, addition and multiplication of images, relabeling of segmented images, connected components labeling, and binary morphological operations (erosion, dilation, opening and closing). All of them are implemented using standard filters of ITK.
4.2
Filtering
There are several classic filtering methods implemented in UsimagTool such as Gaussian, median, bilateral, or anisotropic filters. Additionally we have developed three new filters specific for US images: a Speckle Reduc-ing Anisotropic filter (SRAD) [33], a Tensor guided Anisotropic Diffusion filter (TAD) based on [31] and an anisotropic Wiener filter for speckle removal based on [18].
4.2.1 Speckle Reducing Anisotropic Diffusion (SRAD)
Anisotropic diffusion is a well known and widely used tech-nique in image processing, since it offers a good compro-mise between noise removal and edge preservation; the original formulation by Perona and Malik [22] was based on the heat diffusion equation, where the gray level of the image, I, acts as the temperature that drives the heat con-duction over the image. The diffusion equation follows:
∂I
∂t =∇ ·(c(k∇Ik)· ∇I) (1)
wherec(x)is the diffusion parameter; the higher the value ofc, the faster the diffusion. In the particular case thatcis a constant, eq. (1) can be rewritten as∂I/∂t=c∇2I, with
∇2the Laplacian operator, whose solution is the
convolu-tion of the initial state of the image,I(x,0), with a Gaussian kernel whose standard deviation grows witht. Obviously, as time passes by, the smoothing is greater, so noise is re-moved, but edges are not preserved at all.
The solution is to design a diffusion parameter that avoids the smoothing of edges. The function c(x) is in-tended to decrease as the value ofk∇Ikincreases, so that near the image contours, where the gradient magnitude is large, diffusion is stopped, meanwhile in homogeneous re-gions the smoothing is more severe. The most usual forms forc(x)are:
c(x) = 1
1 + (x/k)2 (2) c(x) = exp −(x/k)2
(3) where kis a parameter that needs to be estimated, see [3] for details. Once c(x) is defined, the partial differential equation (PDE) in eq. (1) is solved with a finite difference scheme.
Although this approach is very useful in the case of addi-tive noise, this may not be the case with US imaging, where this model cannot be assumed. We use here the ideas pro-posed in [33], substituting this gradient-driven diffusion co-efficient with a coco-efficient based on adaptive filtering the-ory; the coefficient of variation used in adaptive filtering
for speckle removal [13, 9] is casted into the PDE environ-ment writing it in terms of the discrete approximations to the derivatives ofI, say ∇I and∇2I, and so an
“instan-taneous coefficient of variation”,q, is obtained that can be seen as a discrete version of the following expression:
q2(x, y, t) = 1 2 k∇Ik I 2 −161 ∇2I I 2 1 +1 4 ∇2I I 2 (4)
And then the diffusion coefficient is estimated as fol-lows:
c(q) = 1
1 + (q2−q2
0)/(q20(1 +q20))
(5) whereq0 ≡q0(t)is a parameter that needs to be iteratively
estimated.
Although the formulation of SRAD is deduced in [33] only for the 2D case, we have extended it to a 3D case. Moreover, given the templated nature of the ITK software, the same source code, with no modifications, is valid for both the 2D and 3D scenarios. A filtered image using this method is shown in figure 5, using as input the image shown in the left viewer of figure 3.
4.2.2 Tensor Guided Anisotropic Diffusion Filtering
(TAD)
One of the main issues with anisotropic diffusion filtering is the numeric scheme used to solve the PDE given by eq. (1). The most usual way is the explicit implementation by means of finite differences: at a given iteration, the right hand side of the equation is computed given the current estimates of
I(x, y, t), and then the partial derivative in time is approx-imated by∂I/∂t ≃(I(x, y, t+τ)−I(x, y, t))/τ so that
I(x, y, t+τ)can be easily found by forward differences as:
I(x, y, t+τ) =I(x, y, t)+τ∇·(c(I(x, y, t))· ∇I(x, y, t))
(6) The problem with this scheme is that it needs a very low value ofτ(typicallyτ <1/4in the 2D case) for the system to be stable, so many iterations are needed to converge to the steady state. The solution given in [31] is to use a semi-implicit scheme, which has been proved to unconditionally converge for any value ofτ; the expression is as follows:
I(x, y, t+τ) =I(x, y, t)+
τ∇ ·(c(I(x, y, t))· ∇I(x, y, t+τ)) (7)
The method is said to be semi-implicit because one uses the gradient in the next time sample but the diffusion co-efficient is assumed to be the one computed in the current
iteration. Each step of the algorithm requires to solve a lin-ear system of equations which is highly sparse, so a recur-sive algorithm may do this task with low computational cost [31].
On the other hand, traditional anisotropic diffusion is based on the fact that diffusion is simply stopped near rel-evant structures, so edges are preserved, but not enhanced. The approach of [1] generalizes the heat diffusion equation by the introduction of a diffusion tensor instead of the diffu-sion parameter, soc(x)is a2×2matrix (in the 2D case); the eigenvectors of this matrix are set to be the same as those of the structure matrix of the image at this point. Regarding the eigenvalues, the one corresponding to the direction par-allel to the contour is kept constant, meanwhile an expres-sion similar to that of eq. (3) is used to compute the eigen-value along the orthogonal direction. This way, near edge pixels there still exists a diffusion process, but it happens in a direction parallel to the contour. Far from the edges, the diffusion is isotropic to favor noise removal. This kind of filters not only preserve the contours, but they are able to enhance them as well. A semi-implicit implementation is derived once again to obtain an adequate computational behavior.
Although the work in [1] is intended for speckle removal, the problem is the same as in the case of scalar anisotropic diffusion: speckle is a multiplicative noise, so the additive model cannot be assumed. For this reason, we have adapted the algorithm to overcome the problem in a similar way to that introduced in [33] for the scalar case. The diffusion tensor is here computed based on local statistics specific for the US imaging case. The direction of the eigenvector par-allel to the contour is found using the sticks model proposed by Czerwinski [6].
For the eigenvalues, a simple speckle classifier [23, 18] based on the moments of the homodynedk-distribution is used to discriminate fully formed speckle, where the dif-fusion is purely isotropic, and resolved speckle, where the diffusion is driven only following the sticks direction, which are ideally parallel to the contours of the US image. In the middle viewer of figure 3 we show a result using this filter, using as input the image in the left viewer.
4.2.3 Anisotropic Wiener Filter with Bias Correction
The Wiener filter, that is, the optimum linear filter in the minimum mean square error (MMSE) allows to estimate a signalXfrom its noisy observationY=X+N, whereN
is the additive noise. If we assume that the data elements are mutually independent and the noise is stationary and its mean is zero, we can formulate the local Wiener filter as [14]. Z(n) = σ 2 X(n) σ2 X(n) +σN2 [Y(n)−ηX(n)] +ηX(n) (8) whereY(n)is then-th element ofY,ηX(n)andσ2X(n) the mean and variance of the signal andσ2
N the variance of the noise. These statistical parameters are actually unknown and must be estimated from the observed data. Under lo-cal ergodicity assumptions, they can be estimated by spatial averages in a neighborhood of each elementn, according to eq. (9) and (10): ηX(n)≈ηY(n)≈ 1 |δ(n)| X m∈δ(n) Y(m) (9) σ2X(n)≈σY2(n)≈ 1 |δ(n)| −1 X m∈δ(n) (Y(m)−ηY(n))2 (10) whereδ(n)is a neighborhood of then-th element. This is a valid estimation in homogeneous regions, but it does not take into account the boundaries or discontinuities in the im-ages. For this reason, the Wiener filter that we have imple-mented, includes the possibility of usekanisotropic neigh-borhoods for the mean and variance estimation. The esti-mated parameters will be the ones computed in the neigh-borhood that aligns the data orientation. It can be assessed as the one where the variance achieves its minimum value.
The estimation of the noise variance is made by eq. (11), where λ ∈ [0,1]is a free parameter, andσmin andσavg are the minimum and mean values, respectively, of the local variancesσY(n).
σ2N =λσmin2 + (1−λ)σavg2 (11)
On the other hand, it deserves to be highlighted that Wiener filter is known to be optimal under the tion of Gaussian distributed data. However, this assump-tion is not valid for US images, that are better modeled by a Rice distribution [29]. For this reason, the direct appli-cation of Wiener filtering would introduce a bias in the re-sults. Therefore, our Wiener filter implementation includes a previous bias correction step based on the Rice distribu-tion. It is based on the fact that the SNR (Signal to Noise Ratio) for a Rice distribution is a monotonic function that depends only on one parameter γ = s
σ [18], where S is the signal envelope and σits variance. The SNR can be estimated from the observation by the method of the mo-ments. Since the SNR function is monotonic, we can es-timate the parameter γ from the estimated SNR value by means of the inverse function. On the other hand, the second order moment of the Rice distribution is given by
EY2 = 2σ2+s2 = s2 2+γ2
γ2 . Therefore, the signal enve-lope can be estimated by the eq. (12).
¯ s= s ¯ Y2γ¯2 2 + ¯γ2 (12)
In such a way, we compute the bias corrected signal
Y′(n) =Y(n)−Y¯(n)+¯s(n)and the anisotropic Wiener
fil-tering is applied over it. The resulting approach eliminates the bias, maintains the underlying signal structure unaltered and at the same time reduces the local variability palliating the speckle effect. In the right viewer of figure 3 we show a result using this filter, using as input the image in the left viewer.
4.3
Segmentation
We include in UsimagTool some classic algorithms such as basic thresholding and level sets. However we will de-scribe here a semiautomatic method of segmentation of the kidney external surface in US images. The main objectives of this task are to improve the common measurements of the renal volume and the visualization and analysis of the organ shape. In the manual part the practitioner inserts six points that correspond to the main axes of the best-fit el-lipsoid for the kidney surface [4]. This simple insertion is extended to those cases in which the whole kidney volume is not acquired, where the user must introduce twelve points that correspond, to the main axes of the visually estimated fit of an ellipse to the extremal and central kidney slice con-tours.
The automatic method is a model based segmentation one. It is a 3D extension of the method in [17]. It takes the points and builds an ellipsoid. Then it discretizes this sur-face by means of a cylindrical coordinate system induced parametrization. So the initial surface is described by a set of S = P ×J points that correspond respectively to the set of cuts and rays of the model. This template surface is refined by allowing its deformation into a discrete space defined by the setΩ = Λ×S, where Λrefers to a set of possible radial deformations in the surroundings of the tem-plate surface.
Each of the elements inΩdefine two regions in the vol-ume image, namely a region of closeness and an inner re-gion. The algorithm is going to construct the two exter-nal potential terms with the pixels included in these two regions. In our framework these terms contribute to the a posteriori likelihood terms. They are based respectively on an exponential distribution modeling of the image gradient and a beta distribution modeling of the image intensity. On the other side, from the locations of the possible deforma-tion points and their neighbors, we can construct a prede-fined a priori model or the internal energy of the model that tries to maintain the smoothness of the surface by favoring those solutions in which the first and second finite radial deformation derivatives are small. Both trends are joined
Figure 6. 2D kidney segmentation
together into a Markov Random Field of deformations. The estimation of the relative importance of each of them as well as the estimation of the deformed surface are integrated in a Maximum a Posteriori framework [18, 12].
A result of a segmented contour in one slice of a healthy kidney is shown in figure 6. The code of the module is based on VTK and it is integrated with the rest of the functionali-ties in UsimagTool, although it is at this time a beta version. It seems clear for us that the method could be adapted for the segmentation of another abdominal organs as the liver or the prostate as pointed out in [17]. This could be interesting for taking a better advantage in the visual interpretation of 3D images acquired with modern echographic techniques.
4.4
Registration
The registration algorithm included in UsimagTool is an adaptation of the “demons” algorithm introduced by Thirion [28], because it is suitable for intra modal de-formable registration. The idea is to consider the iso-intensity contours of one image as semipermeable mem-branes and to deform the other image by pushing it in the normal direction of the contours and diffusing it through those membranes by the action of some effectors in the membranes. The displacement field is derived from the op-tical flow equation:
D(x)∇f(x) =−(m(x)−f(x)) (13) wheref(x)is the fixed image andm(x)is the moving im-age, andD(x)is the displacement field or optical flow be-tween both images. However this equation is not sufficient to obtain locallyD(x)and some form of regularization as well as re-normalization is used. The final displacement is obtained iteratively. Given the initial deformation,D0(x),
the displacement in then-th iteration will be:
Dn(x) =Dn−1(x)−(m(x+D
n−1(x))−f(x))∇f(x)
||∇f||2+ (m(x)−f(x))2
(14) The problem with the classic demons algorithm is that it cannot cope with variations in the gray level, even if they
are very small, for example the intensity bias present in MRI imaging. In [5] this problem is overcome deriving a generic expression from eq. (14):
Dn(x) =Dn−1(x)− 2E(x)
k∇E(x)k2+ 4λE(x)∇E(x) (15)
where the update rule forD(x)is obtained as a function of ∇E(x), withE(x)a locally defined similarity measure or similarity energy. In [5] this energy is assumed to be the lo-cal correlation between both images to register, so that their corresponding intensities have no longer to be the same, but instead it is enough that they follow a linear relation.
Besides, the computation of the local energy, as well as its derivatives, may be easily done by means of convolutions with a set of Gaussian kernels, with low computational load if a recursive IIR filtering is used instead of the explicit con-volutions [7].
Given that registration is an ill posed problem, in order to force a realistic deformation field, we can use an elastic or fluid regularization.
We have adopted here the approach in [5], together with a regularization step (elastic, fluid, or both types simultane-ously are allowed) based on Gaussian smoothing and a mul-tiresolution strategy, as a versatile, accurate, and robust al-gorithm for elastic registration. We show the moving, fixed and registered images of a simulated US data of a human brain in figure 7.
5
Conclusions and future work
We have presented here a new tool for the development and use of algorithms for US image processing, that can be also used for other image modalities. The features of this tool makes it quite useful for researchers, due to the flexibil-ity offered by its simple architecture, that allows to include new algorithms very fast. It is also important to highlight that this is an open source project which is available on line in http://www.lpi.tel.uva.es/usimagtool.
We have also presented five algorithms for US im-age processing: three filter algorithms, one segmentation method and one registration method, whose implementa-tions are completely new, based on well known published works, and developed as ITK and VTK classes, but that are not part of the official ITK and VTK release. To the best of our knowledge the SRAD filter offered here is the first im-plementation that extends this filter to 3D and the combina-tion of the methods proposed in [1],[6] and [23] in the TAD filter is a novel approach. For these reasons we state that UsimagTool is the most advanced software package for US image processing available currently for the research com-munity.
Future work includes several usability tests to improve user interaction, improving the robustness of the GUI and include more image processing algorithms not only for US images, as well as more basic image operations support.
Acknowledgments
This work has been funded by the Spanish Ministry of Science and Technology grant TEC-2004-06647-C03-01, and has been partially supported by an European co-funded grant FP6-507609 from the European Commission.
References
[1] K. Z. Abd-Elmoniem, A. M. Youssef, and Y. M. Kadah. Real-time speckle reduction and coherence enhancement in ultrasound imaging via nonlinear anisotropic diffusion.
IEEE Trans. on Biomedical Engineering, 49(9):997–1014,
2002.
[2] Able Software Corp. 3D Doctor,
http://www.ablesw.com/3d-doctor.
[3] S. Aja-Fernandez and C. Alberola-Lopez. On the estima-tion of the coefficient of variaestima-tion for anisotropic diffusion speckle filtering. IEEE Transactions on Image Processing, 15(9):2694–2701, 2006.
[4] J. Bakker, M. Olree, R. Kaatee, E. de Lange, K. Moons, J. Beutler, and F. J. Beek. Renal volume measurements: accuracy and repeatability of us compared with that of mr imaging. Radiology, 211:623–628, 1999.
[5] P. Cachier and X. Pennec. 3D non-rigid registration by gra-dient descent on a Gaussian-windowed similarity measure using convolutions. In Procs. of MMBIA ’00, pages 182– 189, Hilton Head Island, South Carolina, USA, June 2000. IEEE Computer society.
[6] R. N. Czerwinski, D. L. Jones, and W. J. O’Brien. Line and boundary detection in speckle images. IEEE Trans. on
Image Processing, 7(12):1700–1714, 1998.
[7] R. Deriche. Fast algorithms for low-level vision. IEEE Trans. on Pattern Analysis and Machine Intelligence,
12(1):78–87, 1990.
[8] P. Fillard, N. Toussaint, and X. Pennec. MedINRIA: DT-MRI processing and visualization software, November 2006. Guest paper at the Similar Tensor Workshop, Las Pal-mas, Spain.
[9] V. S. Frost, J. A. Stiles, K. S. Shanmugan, and J. C. Holtz-man. A model for radar images and its application to adap-tive filtering of multiplicaadap-tive noise. IEEE Trans. on Pattern
Analysis and Machine Intelligence, 4(1):157–165, 1982.
[10] D. Gering, A. Nabavi, R. Kikinis, W. E. L. Grimson, N. Hata, P. Everett, F. Jolesz, and W. W. III. An integrated visualization system for surgical planning and guidance us-ing fusion and interventional imagus-ing. In MICCAI, pages 808–819, Cambridge, England, Sept 1999.
[11] L. Ibanez, W. Schroeder, L. Ng, and J. Cates. The ITK Software Guide. Kitware, Inc. ISBN 1-930934-15-7, http://www.itk.org/ItkSoftwareGuide.pdf, second edition, 2005.
Figure 7. Registration in UsimagTool
[12] S. Lakshmanan and H. Derin. Simultaneous parameter esti-mation and segmentation of gibbs random fields using simu-lated annealing. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 11(8):799–813, Aug. 1989.
[13] J. S. Lee. Digital image enhancement and noise filtering by using local statistics. IEEE Trans. on Pattern Analysis and
Machine Intelligence, 2(1), 1980.
[14] J. Lim. Two Dimensional Signal and Image Processing.
Prentice Hall, 1990.
[15] A. Loening and S. Gambhir. Amide: a free software tool for multimodality medical image analysis. Molecular Imaging, 2(3):131–137, 2003.
[16] Y. Lu, S. Pathak, L. Gong, J. Goldschneider, and P. Kinahan. Open source medical image fusion display.
[17] M. Martin and C. Alberola. An approach for contour detec-tion of human kidneys from ultrasound images using markov random fields and active contours. Medical Image Analysis, 9(1):1–23, 2005.
[18] M. Martin-Fernandez, E. Mu˜noz-Moreno, and C. Alberola-Lopez. A speckle removal filter based on an anisotropic wiener filter and the rice distribution. In IEEE
Interna-tional Ultrasonics Symposium, pages 1694–1697,
Vancou-ver, Canad, Oct 2006.
[19] Mayo Clinic. ANALYZE software, http://www.mayo.edu/bir/Software/Analyze/Analyze.html.
[20] Mercury Computer Systems Inc. Amira , http://www.amiravis.com.
[21] MIT Artificial Intelligence Lab and the Surgical Planning Lab at Brigham & Women’s Hospital, Harvard Medical School. 3D Slicer. 3D Slicer software, http:// www.slicer.org. [22] P. Perona and J. Malik. Scale space and edge detection using anisotropic diffusion. IEEE Trans. on Pattern Analysis and
Machine Intelligence, 12(1):629–639, 1990.
[23] R. W. Prager, A. H. Gee, G. M. Treece, and L. H. Berman. Decompression and speckle detection for ultrasound images using the homodyned k-distribution. Pattern Recognition
Letters, 24:705–713, 2003.
[24] Research center caesar, Bonn, Germany. The Julius software
platform, http://www.julius.caesar.de.
[25] A. Rosset, L. Spadola, and O. Ratib. Osirix: An open-source software for navigating in multidimensional dicom images.
J Digit Imaging, 17(3):205–216, Sep 2004.
[26] W. Schroeder, K. Martin, and W. Lorensen. The
Visual-ization Toolkit: An Object Oriented Approach to Computer Graphics. Kitware, Inc, third edition, 2004.
[27] M. Sweet, C. Earls, and B. Sptizak. FLTK Fast Light Toolkit
1.1.7, programming manual, 2006.
[28] J.-P. Thirion. Image matching as a diffusion process: an analogy with maxwell’s demons. Medical Image Analysis, 2(3):243–260, 1998.
[29] T. A. Tuthill, R. H. Sperry, and K. J. Parker. Deviation from rayleigh statistics in ultrasonic speckle. Ultrasonic Imaging, 10:81–89, 1988.
[30] V. Nicolas, UCL. Medical Studio http://www.medicalstudio.org.
[31] J. Weickert, B. ter Haar Romery, and M. Viergever. Ef-ficient and reliable schemes for nonlinear diffusion filter-ing. IEEE Transactions on Image Processing, 7(3):398–410, march 1998.
[32] I. Wolf, M. Vetter, I. Wegner, T. Bttger, M. Nolden, M. Schbinger, M. Hastenteufel, T. Kunert, and H. Meinzer. The medical imaging interaction toolkit. Medical Image Analysis, 9:594–604, 2005.
[33] Y. Yongjian and S. Acton. Speckle reducing anisotropic diffusion. IEEE Transactions on Image Processing,
11(11):1260–1270, Nov 2002.
[34] P. Yushkevich, J. Piven, H. Hazlett, R. Smith, S. Ho, J. Gee, and G. Gerig. User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability. Neuroimage, 2006. To appear.
