当前位置:首页 >> IT认证 >>

Nondistorting flattening maps and the 3D visualization of colon CT images


Nondistorting Flattening Maps and the 3D Visualization of Colon CT Images
Steven Haker Department of Diagnostic Radiology Yale University New Haven, Connecticut 06520 Sigurd Angenent Department of Mathematics University of Wisconsin Madison, Wisconsin 53705 Allen Tannenbaum? Department of Electrical and Computer Engineering Georgia Institute of Technology Atlanta, GA 30332-0250 email: tannenba@ece.gatech.edu fax: 404-894-7995 Ron Kikinis, M. D. Harvard Medical School Brigham and Women’s Hospital Harvard University Boston, MA 02115 September 18, 2000

1

This work was supported in part by grants from the National Science Foundation DMS-9058492, ECS-99700588, NSF-LIS, by the Air Force Of?ce of AF/F49620-981-0168, by the Army Research Of?ce DAAG55-98-1-0169, and MURI Grant. ? All correspondence concerning this paper should be sent to Allen Tannenbaum.
Abstract In this paper, we consider a novel 3D visualization technique based on surface ?attening for virtual colonoscopy. Such visualization methods could be important in virtual colonoscopy since they have the potential for non-invasively determining the presence of polyps and other pathologies. Further, we demonstrate a method which presents a surface scan of the entire colon as a cine, and affords viewer the opportunity to examine each point on the surface without distortion. We use certain angle-preserving mappings from differential geometry in order to derive an explicit method for ?attening surfaces obtained from 3D colon CT imagery. Indeed, we describe a general method based on a discretization of the Laplace-Beltrami operator for ?attening a surface onto the plane in a manner which preserves the local geometry. From a triangulated surface representation of the colon, we indicate how the procedure may be implemented using a ?nite element technique, which takes into account special boundary conditions. We also provide simple formulas which may be used in a real time cine to correct for distortion. Key words: Virtual colonoscopy, CT colonography, ?attening maps, shape preservation. This paper has been accepted for publication to IEEE Transactions on Biomedical Engineering.

1 Introduction
3D visualization is becoming an increasingly important technique in surgical planning, non-invasive diagnosis and treatment, and image-guided surgery. Surface warping and ?attening, which allow the easy visualization of highly undulated surfaces, are methods that are becoming increasingly widespread. For example, ?attened representations of the brain cortical surface are essential in functional magnetic resonance imaging since one wants to show neural activity deep within the folds or sulci of the brain. 3D visualization is also of great importance in virtual colonoscopy in which one can non-invasively determine the presence of pathologies. Virtual colonoscopy is currently an active area of research by radiologists as a minimally invasive screening method for the detection of small polyps (see [5, 6] and the references therein). In the colon, this has become possible because of imaging devices which allow single breath hold acquisitions of the entire abdomen at acceptable resolutions. Most reports have focused on methods which use computer graphics to simulate conventional colonoscopic procedures [6, 11, 14]. Other relevant references include McFarland,Valev. Virtual colonoscopy has some fundamental problems, which it shares with conventional colonoscopy. The most important one is that the navigation using inner views is very challenging and it happens frequently that sizable areas are not inspected at 2

all, leading to incomplete examinations. An alternative approach for the inspection of the entire surface of the colon is to simulate the approach favored by pathologists, which involves cutting open the tube represented by the colon, and laying it out ?at for a comprehensive inspection. In some very recent work [9], a visualization technique is proposed using cylindrical and planar map projections. It is well-known that such projections can cause distortions in shape as is discussed in [9] and the references therein. In this paper, we take another approach. We present a method for mapping the colon onto a ?at surface in a conformal manner. A conformal mapping is a one-to-one mapping between surfaces which preserves angles, and thus preserves the local geometry as well. Our approach to ?attening such a surface is based on a certain mathematical technique from Riemann surface theory, which allows us to map any highly undulating tubular surface without handles or self-intersections onto a planar rectangle in a conformal manner. There is some related work in the interesting paper [16] on the topological ?attening of a tube onto the plane and its application to virtual colonoscopy. In [16], an electro-magnetic ?eld is simulated by placing charges along a ?y-through path. The resulting ?eld lines which emanate from a point on the path de?ne a surface whose intersection with the colon surface forms a loop which is ?attened into the plane. Our approach differs in that no ?ight path needs to be calculated and the conformal nature of the ?attening allows us to easily correct for distortion. From a triangulated surface representation of the colon, we indicate how the procedure may be implemented using ?nite elements. Moreover, we explicitly show how various structures of the colon may be studied using this approach. In contrast to virtual colonoscopy methods which are based on a “?y-through,” our method allows the entire colon surface to be viewed at once, unobscured by surface folds. Hence the ?attening mapping can be used to obtain an atlas of the given surface in a straightforward, canonical manner. The ?attening function may be obtained as the solution of a second order elliptic partial differential equation (PDE) on the surface to be ?attened. From a triangulated representation of the colon surface, we use a ?nite element procedure to numerically approximate the ?attening function. This numerical method involves mearly solving two sparse systems of linear equations and can be accomplished by using standard conjugate gradient techniques. Once the colon surface has been ?attened onto a rectangular region of the plane, we utilize a method by which the entire colon is presented as a cine, and which allows the viewer to examine each surface point without distortion at some time in the image sequence. Thus in this sense we achieve 100% view with 0% distortion. In this present paper, we are interested in exploiting our conformal methodology for virtual colonoscopy. However, in addition to the colon, we have been employing our method in the study of depth maps of bladder images, and in the study of curvatures and polar maps from ultrasound heart images. We are also interested in using our work for the construction of canonical brain models and coordinates on both the grey matter and grey/white matter boundaries from 3D MR data sets [1]. We now summarize the contents of this paper. In Section 2, we give a high-level overview of our ?attening method. In Section 3, we describe the approximation method we employ for numerically computing the ?attening mapping based on ?nite elements. In Section 4, we provide formulas which can be used in real time to create a cine which 3

provides a view of each point of the surface without distortion. Then in Section 5, we illustrate the methodology on some CT colon imagery, and in Section 6, we draw some conclusions about our general approach as well as directions for future research. Finally in Section 7, we provide some key mathematical details justifying our methodology.

2 General Approach to Colon Flattening
? We ?rst consider a mathematical model for the colon surface. Accordingly, let represent an embedded surface (no self-intersections) which is topologically an openended cylinder. In the mathematical appendix (see Section 7 below), we will give more details on the analytical basis for ?attening such a surface onto the plane. In this section, we just sketch some of the key points. We assume that is a smooth manifold. For the ?nite element method described in the next section, it will be enough to take it as a triangulated surface. We refer the reader to [3] for the basic theory of surfaces of the type we are considering here, and to [10] for the relevant results on partial differential equations. The boundary of consists of two topological circles, which we will call ? and ? We want to construct a conformal map (i.e., a map which preserves angles and is one-to-one), , which sends to an annulus such that ? and ? are mapped to the inner and outer boundary circles of respectively. The construction of begins with ?nding a solution ? to the Dirichlet problem ? on ? ? on ? and ? on ? ? with boundary conditions ? Here is the Laplace-Beltrami operator on the surface (see [10]). We then ?nd a smooth curve on which runs from ? to ? such that ? is strictly increasing along . This curve de?nes a cut on , and the cut surface ? is conformally equivalent to a rectangle in the plane. We next compute the harmonic function ? which is conjugate to ? by specifying boundary conditions on the cut surface and again solving a Dirichlet problem. The details for this are given in Section 7. The mapping ? ? sends the surface ?· ? sends to a rectangle, its exponential to an annulus.

? ?

?

?

?

?

?

? ?? ?

?

?

? ?

?

?

?

?

?

·

?

3 Approximation of the Flattening Function
In the previous section, we outlined the analytical procedure for ?nding the ?attening map . Here we will discuss the ?nite element method for ?nding an approximation to this mapping. See [7] for details about this method. Since the procedure given here is crucial for the ?attening algorithm we have used in our simulations, we will describe this implementation in some detail. In [1], we described a related method for brain ?attening. However, because of the differences in topology between the brain and colon surface, the boundary conditions for the ?attening maps are quite different. We assume that is a triangulated surface, and we look for a ?attening map which is continuous on and linear on each triangle. Here, we will concentrate on ?nding ?, the method for its conjugate ? being similar.

?

?

4

It is a classical result [10] that the harmonic function Dirichlet functional

? is the minimizer of the

???
?
?

? ?

?

?

?? ? ?
?

?

? ? ?

where ?? is the gradient with respect to the induced metric on . Let ? ? denote the ?nite dimensional space of piecewise linear functions on . For each vertex ? ? , let ? be the continuous function such that

???

? ? ?? ? ? ?? ?
?

is linear on each triangle.

? ?

?

?

(1)

Now this set as

?

forms a basis for ? ?

???, and so any ? ? ? ???? can be written ?
??
?

?

?

vertex of

So to approximate the solution to the PDE, we minimize which satisfy the boundary conditions. To minimize ? , we introduce the matrix

??? over all ? ? ? ????

??

??

? ? ??

?

?

for any pair of vertices ? ? It is easy to see that ? ? if and only if ? ? are con? nected by some edge in the triangulation. One can show (see [7]) that ? ? ?? ? minimizes the Dirichlet functional over ? ? with the boundary conditions, if for each vertex ? ? ? ? ?,

?

??

?

???

? ???? ?

??

??

??

?

?? ?? ?

(2)

3.1 Computing

Assume ? ? ? ? ? , then the edge ? ? belongs to two triangles, ? If ? ? ? , and ? ? . An essential formula from ?nite-element theory [7], then says that
??

??

??

?

? ? ?? ?

· ??
, and is the angle at

where is the angle at the vertex the vertex in the triangle ? ? . ? , then one has If ?
??

in the triangle ? ?

?

?? ? ?

5

3.2 Summary of Method
We may summarize the ?nite element procedure for the construction of the ?attening map as follows: (1) Compute the
??

(2) Solve the linear equation (2) to obtain the piecewise linear harmonic function ?

?

?

??

?

(3) Cut the surface from ? to ? and compute the boundary values of ? by inte? grating ? (the derivative of ? in the normal direction [10]) around the boundary. Use ?nite elements to solve for ? To ?nd a cut, one may start at a vertex on ? and move from each vertex to the adjacent vertex which has the largest value of ? associated with it. The maximum principle implies that there is always an adjacent vertex with a larger value of ? than the current vertex. Also, in this way we make a cut which follows closely the gradient of ? while remaining on edges of triangles. Thus our ?attened image will be roughly rectangular in shape.

4 Inspection and Distortion Removal
In practice, once the colon surface has been ?attened into a rectangular shape, it will need to be visually inspected for various structures. In this section, we present a simple technique by which the entire colon surface can be presented to the viewer as a sequence of images or cine. In addition, this method allows the viewer to examine each surface point without distortion at some time in the cine. Here, we will say a mapping is without distortion at a point if it preserves the intrinsic distance there. It is well known that a surface can not in general be ?attened onto the plane without some distortion somewhere (see [4]). However, it may be possible to achieve a surface ?attening which is free of distortion along some curve. A simple example of this is the familiar Mercator projection of the earth, in which the equator appears without distortion. See [9] for a nice discussion of the classical geographic projections and their application to virtual colonoscopy. In our case, the distortion free curve will be a level set of the harmonic function ? described above (essentially a loop around the tubular colon surface), and will correspond to the vertical line through the center of a frame in the cine. This line is orthogonal to the “path of ?ight” so that every point of the colon surface is exhibited at some time without distortion. Speci?cally, suppose we have conformally ?attened the colon surface onto a rect?? ? ? ? . Let be the inverse of this mapping, and let ? angle ? ? ? ? be the amount by which scales a small area near ? ? , i.e. let be the “conformal factor” for . Fix ? , and for each ?? ? ?? ? de?ne a subset ?? ?? ? ? ?? ? ? ? ? which will correspond to the contents of a cine frame. We de?ne a mapping

? ? ?

?

·

?

?

? ? ?

?

?? ??

?? ??

? ??

? ??
6

?

?

??? ?

from the formula for that lengths measured in the ? direction accurately re?ect the lengths of corresponding curves on the colon surface.

?? ?? ???? ? ? ?? ?? ?? ? ??? ?? and in particular ??? ?? ??? ?? ? ? ? The conformality of the ? ? ??? ??, implies that the composition of ?attening map, together with this value for the ?attening map with sends the level set loop ? ?? on the colon surface to ? in the ?–? plane without distortion. In addition, it follows the vertical line ?
?? ?? ?? ??

from ?? to

which has differential

5 Computer Simulations
We tested our algorithm on three different data sets provided to us by the Surgical Planning Laboratory of Brigham and Women’s Hospital. Since the results were similar in each case, we just include here one of the sets. This example consists in a test of ?attening the colon surface contained in a ? ? CT colon image. Four slices from this data set are shown in Figure 1. First, using the fast segmentation methods of [8, 13] we found the colon surface. Unfortunately, the segmentation algorithm itself does not guarantee that the surface found will be a topological cylinder. In fact, it may contain numerous minute handles which arise because the boundary of the colon, as represented in the data set, may not be sharp. We used a morphological based method by which these handles can be effectively removed and a surface which has the the topology of a closed-ended cylinder can be extracted. This is done in such a way that the large-scale geometry of the surface is not adversely affected. We then created a triangularization of this surface using the Visualization Toolkit [12]. Next, the triangularized surface was smoothed slightly to reduce the effects of aliasing. This was done by using the ?ow according to mean curvature. This also allowed us to obtain a measure of the convexity and concavity of points on the colon surface by considering the mean curvature vector. The triangles making up the two ends of the tubular surface were then removed to produce an open-ended tube with two boundary curves. Next, as described in the Section 3, we made a cut along the colon surface from a point on one of the bounding curves to a point on the other bounding curve, and then conformally ?attened the resulting cut surface onto a nearly rectangular region of the plane. We colored points of the colon surface according to mean curvature, and then applied these same colors to the ?attened image in the plane. This gave an elegant way to visualize the structure of the colon which is a twisted, undulating tube. We should note that the whole procedure described above from segmenting the colon to ?attening took about 12 minutes on a Sun Ultrasparc 10 Workstation. In Figure 2, we show three different views of the segmented colon surface and its ?attened representation. A small circled region visible at the bottom of the ?attened surface, and the corresponding region of the original surface, is shown in detail in

?

?

??

7

Figure 3. In Figure 4, we show an interior “?y-through” view of the colon image, and its ?attened representation. Notice how much easier it is to get a view of the whole interior region in this manner. Next, we used the formulas in Section 4 to create a distortion correcting cine of a scan of the colon surface. To make numeric integration fast and simple, we interpolated the function (de?ned on the rectangular representation of the colon in the ?–? plane) onto an equally spaced grid. For each particular value ?? of ?, increasing from to ?? ? on this grid, we mapped the corresponding sub-rectangle ?? to the ?–? plane as described in Section 4. We found that this could easily be done in real time. Four frames from the resulting cine can be seen in Figure 5. For comparison, we show an exterior view of the portion of the original colon surface corresponding to each frame. As the cine progresses, the vertical line through the center of the frame is a distortion correcting representation of a loop on the colon surface. These loops sweep continuously over the colon surface from end to end, and thus every surface point is presented to the viewer in some frame without distortion. Further, the mapping to the ?–? plane is a continuous function of time, and so there is no jumping between frames. The fact that distances in the ? direction are the same as corresponding distances on the surface seems to help in keeping distortion to an acceptable level even off the vertical center line. Although these images show only mean curvature, we believe that other geometric quantities may also be useful in the visualization of the ?attened surface. Polyps, for example, should have relatively high Gaussian curvature as compared to the ?atter surrounding colon surface. It is possible that thickness of the colon wall may also be used to distinguish these pathologies.

?

6 Conclusions
In this note, we outlined a high-level procedure based on conformal geometry and harmonic analysis for the construction of a ?attening map of a colon surface derived from volumetric CT data. Further, we presented a numerical algorithm which ?nds this map based on ?nite elements. Finally, we demonstrated formulas which can be used real time in a cine to provide a distortion free view of every point on the colon surface. Additional details on the underlying mathematics, numerics, and an application to 3D brain imagery may be found in [1]. Deformations of highly undulated surfaces have many uses for both military and medical imagery, as well as computer graphics in the area of texture mappings. In this paper, we have proposed a novel solution to the problem of surface deformation in a manner in which angles are locally preserved. We have also been considering area-preserving ?attening maps of minimal distortion, that is, to try to optimize the trade-off of exact area-preservation with minimal angle deviation in our ?attening maps. It is mathematically impossible to have both the area and angles preserved everywhere in a diffeomorphism between surfaces unless the two surfaces have the same Gaussian curvature; see [4] for all the relevant results and de?nitions. Thus the type of trade-off mentioned above is probably the best that one can do. For the mathematical details see [2]. We will be applying this latter 8

methodology to both brain and colon imagery in our future work.

7 Mathematical Appendix
We use the notation of Section 2. We will ?ll in the details of the high-level algorithm how to construct the ?attening map .

?

(1) Let ? be the solution of the Dirichlet problem

? ?

??

Here represents the Laplace-Beltrami operator on the surface . This second order partial differential operator is intrinsic to the surface and is a generalization of the standard Laplacian in the plane. From the standard theory of partial differential equations [10], ? exists and is unique. (2) Let be a smooth curve on which runs from ? to ? such that ? is strictly increasing along . The maximum principle implies the existence of such a curve. This curve de?nes a cut on , and the cut surface ? is conformally equivalent to a rectangle in the plane.

?

? on ??? ? on ? ? on ?

?

??

(3)

?

?

?

?

Let be the oriented boundary of this cut surface, i.e. let run around ? , then along to ? , around ? , then along again in the opposite direction to ? . The closed path must run around ? and ? in a fashion consistent with an orientation given on the surface We want to compute the harmonic function ? which is conjugate to ? by specifying boundary conditions on the cut surface and again solving a Dirichlet problem. Let × and ? denote tangential and normal derivatives respectively. By the Cauchy-Riemann equations, one has

?

on

? ? × ? and so the boundary values of ? may be found by integration along ??

?

?

? × ×

?

? × ?

Note that if we choose so that it follows the gradient of ? from ? to ? , then we would have ? along , and so ? would be constant on along the ? parts corresponding to .

where here × is the arc-length along . Since ? is harmonic, the divergence ? ? theorem guarantees that , and so ? is smooth (periodic) on ? ×

?

?

9

(3) The component ? of the rectangular map ? ? ? is then found by solving the Dirichlet problem using the calculated boundary values for ? . If is chosen to follow along the gradient of ?, then the image of in the ?–? plane under ? will be a rectangle of width 1 and some height . For other curves , along which ? is monotone, the image will be a region in the ?–? plane between the lines ? ,? , the graph of a function ? , and a copy of the graph . The fact that the distance of shifted vertically by a constant, i.e. ? across the region vertically is constant is again a consequence of the divergence theorem.

·

?

?

?

? ?·

??

We may now dilate the ?attened rectangular image of under ? homothetically, We may then map the surface to a true and so we may assume that cylinder of radius 1 by simply “rolling up” the ?attened image. We may also map ? , effectively reconnecting the image of the surface it to an annulus via smoothly across the cut. We may then map it to the sphere using stereographic projection, if desired. In practice, we have found the rectangular mapping to be the most natural representation of the ?attened colon. The existence of the cut in the rectangular mapping does not present a problem for visualization; the constant height of the rectangular image allows us to extend it periodically above and below the cut.

?

?

Remark: The annular map is, any other conformal boundary is given by for some constants ?

?

is unique up to a dilation, rotation and translation. That which maps to an annulus and ? to its outer

?

??

?

·?

(4)

References
[1] S. Angenent, S. Haker, A. Tannenbaum, R. Kikinis, “Laplace-Beltrami operator and brain surface ?attening,” IEEE Trans. Medical Imaging, 18 (1999), pp. 700711.. [2] S. Angenent, S. Haker, A. Tannenbaum, R. Kikinis, “On area preserving mappings of minimal distortion,” in System Theory: Modeling, Analysis, and Control, edited by T. Djaferis and I. Schick, Kluwer, Holland, 1999, pages 275-287. [3] H. Farkas and I. Kra, Riemann Surfaces, Springer-Verlag, New York 1991. [4] M. P. Do Carmo, Riemannian Geometry, Prentice-Hall, Inc. New Jersey, 1992. [5] A. Hara, C. Johnson, J. Reed, “Detection of colorectal polyps by computed tomographic colography: feasibility of a novel technique,” Gastroenterology 110 (1996), pp. 284-290.

10

[6] A. Hara, C. Johnson, J. Reed, R. Ehman, D. Ilstrup, “Colorectal polyp detection with CT colography: two- versus three-dimensional techniques,” Radiology 200 (1996), pp. 49-54. [7] T. Hughes, The Finite Element Method, Prentice-Hall, New Jersey, 1987. [8] S. Kichenasamy, P. Olver, A. Tannenbaum, A. Yezzi, “Conformal curvature ?ows: from phase transitions to active contours,” Archive Rational Mechanics and Analysis 134 (1996), pp. 275-301. [9] D. Paik, C. Beaulieu, R. Jeffrey, C. Karadi, and S. Napel, “Visualization modes for CT colonography using cylindrical and planar map projections,” Technical Report, Department of Radiology, Stanford University School of Medicine, Stanford, CA, 1999. [10] J. Rauch, Partial Differential Equations, Springer-Verlag, New York 1991. [11] G. Rubin, S. Napel, and A. Leung, “Volumetric analysis of volumetric data: achieving a paradigm shift,” Radiology 200 (1996), pp. 312-317. [12] W. Schroeder, H. Martin, B. Lorensen, The Visualization Toolkit, Prentice-Hall, New Jersey, 1996. [13] K. Siddiqi, A. Tannenbaum, and S. Zucker, “Area and length minimizing ?ows for image segmentation,” IEEE Trans. Image Processing 7 (1998), pp. 433-444. [14] D. Vining, “Virtual endoscopy: is it reality?,” Radiology 200 (1996), pp. 30-31. [15] B. Wandell, S. Engel, and H. Hel-Or, “Creating images of the ?attened cortical sheet,” Invest. Opth. and Vis. Sci. 36 (S612), 1996. [16] G. Wang, E. McFarland, B. Brown, and M. Vannier, “GI tract unraveling with curved cross sections,” IEEE Trans. Med. Imaging 17 (1998), pp. 318-322.

11

Figure 1: Four Slices of CT Colon Data

12

13 Figure 2: Three Views of Segmented Colon and Flattened Representation

Figure 3: Region of Interest in Segmented Data and Flattened Representation

Figure 4: “Fly-Through” View and Flattened Representation

14

Figure 5: “Fly-Through” View and Flattened Representation

15

Figure 6: Frames From Distortion-Correcting Cine with Original Surface 16


相关文章:
更多相关标签: