Click here to load reader

Advanced Curved Planar Reformation: Flattening of Vascular …€¦ · Advanced Curved Planar Reformation: Flattening of Vascular Structures ... This requirement is satisfied by

  • View

  • Download

Embed Size (px)

Text of Advanced Curved Planar Reformation: Flattening of Vascular …€¦ · Advanced Curved Planar...

  • Advanced Curved Planar Reformation: Flattening of Vascular Structures

    Armin Kanitsar∗ Rainer Wegenkittl‡ Dominik Fleischmann† Meister Eduard Gröller∗

    ∗ Institute of Computer Graphics and Algorithms † Department of Radiology ‡ TIANI MedgraphVienna University of Technology Stanford University Austria

    Figure 1: An untangled vascular tree of the peripheral arteries.


    Traditional volume visualization techniques may provide incom-plete clinical information needed for many applications in medicalvisualization. Especially in the area of vascular visualization im-portant features such as the patent lumen of a diseased vessel seg-ment may not be visible. Curved Planar Reformation (CPR) hasproven to be an acceptable practical solution. Existing CPR tech-niques, however, still have diagnostically relevant limitations. Inthis paper we introduce two advanced methods for efficient vesselvisualization, based on the concept of CPR. Both methods benefitfrom relaxation of spatial coherence in favor of improved featureperception. We present a new technique to visualize the interior ofa vessel in a single image. A vessel is re-sampled along a spiralaround the vessel central axis. The helical spiral depicts the vesselvolume. Furthermore, a method to display an entire vascular treewithout mutually occluding vessels is presented. Minimal rotationsaround the branching points of a vessel tree eliminate occlusions.For each viewing direction the entire vessel structure is visible.

    CR Categories: J.3.2 [Computer Applications]: Life and Med-ical Sciences—Medical Information Systems I.3.3 [ComputingMethodologies]: Computer Graphics—Picture/Image Generation

    Keywords: computed tomography angiography, vessel analysis,curved planar reformation


    Non-invasive imaging of the vascular system with computed to-mography (CT) and magnetic resonance imaging (MRI) has be-

    ∗{kanitsar,meister},†[email protected]‡{rainer.wegenkittl}

    come a well established alternative to invasive intraarterial angiog-raphy. CT and MRI provide high-resolution volumetric data sets ofthe human body. These data, however, may contain many objectsof less or no diagnostic interest. This makes volume-rendering (i.e.,Maximum Intensity Projection (MIP), ray casting, shaded surfacedisplay) without preprocessing often impossible or inaccurate.

    CPR - Curved Planar Reformationis a way to visualize vascu-lar structures with small diameter. High level information as thevessel’s centerline is used to re-sample and visualize the data. Thewhole length of the tubular structure is displayed within a singleimage by this technique. Vascular abnormalities (i.e., stenoses, oc-clusions, aneurysms and vessel wall calcifications) are then investi-gated by physicians.

    Current CPR techniques allow the investigation of the vessel lu-men in a longitudinal section through the central axis of the vessel.However, vascular abnormalities might not be touched by this planeand therefore they do not appear in the generated image. One wayto overcome this problem is to rotate the re-sampled plane aroundthe central axis. This results in a set of images to be interpreted bythe radiologist. A more comprehensive display of the entire vascu-lar lumen in one representative image is highly desirable. A newvisualization method was developed to overcome this limitation.

    Another important aspect in computedtomography angiography(CTA) is the efficient visualization of treelike vascular structuresusing CPR display techniques. Multi-path CPR techniques basedon a projective combination of vessel segments provide a spatiallycoherent display of the vascular anatomy [6]. However, parts of thearteries might be superimposed by other arteries depending on theintersecting plane. For a detailed inspection of the entire vasculartree different sections through the vessel’s central axis have to be re-sampled. In order to have diagnostically valuable results the vessellumen should be visible within each image. Thus a new techniquefor unobscured displaying of an arterial tree is proposed.

    Section 2 describes related work in this area. In Section 3 a newmethod for visualizing the interior of a vascular structure is pre-sented. A technique for displaying the entire vascular tree withoutoverlapping arteries is presented in Section 4. Possible improve-ments and conclusions are discussed in Section 5.

  • Figure 2: Traditional CPR (a), and helical CPR (b) generation.


    The most important prerequisite for CPR visualization is an appro-priate estimation of the vessel centerline. Latest CT technology,such as multiple detector-array CT, provide high resolution volu-metric data sets. Due to the large size of these data sets (up to 1900transverse cross-sectional images of the abdomen and entire legs),the manual definition of the vessel centerline is no longer an option.In this respect several algorithms [7, 10, 11] have been developedwith different properties concerning reliability, speed and accuracy.

    Avants and Williams presented a vessel tracking method consist-ing of two parts [2]. From user defined seed points a surface expan-sion is computed based on the eikonal partial differential equation.A minimal cost path is calculated from these regions. From thispath a cross-sectional area/radius profile is generated.

    He et al [5] proposed a path extraction method based on a two-dimensional region-growing algorithm with a subsequent shortestpath algorithm. The resulting path was refined using the multi-scalemedial response. The vascular tree is flattened in a semiautomaticmethod calledMedial Axis Reformation.

    Some authors propose to take the central-axis as an input for thegeneration of an abstract vessel-model. Abstract vessel-models al-low fast rendering, as polygonal meshes of low complexity are gen-erated [3]. Furthermore non-photorealistic rendering provides thepossibility to emphasize global properties of the vascular tree [4].

    Kanitsar et al [6] compared three methods for CPR generation:Projected CPR, stretched CPR and straightened CPR. In additionthree extensions of CPR, developed to overcome the most relevantclinical limitations of CPR visualization have been proposed: ThickCPR, rotating CPR and multi-path CPR. The latter provides a dis-play of a whole vascular tree within one image. While superimpo-sition of bones and arteries is prevented, the intersection of arteriesitself is not avoided.

    Further information about the clinical relevance of the CPR vi-sualization technique can be found in [9, 1, 8].


    The basic idea of helical CPR visualization is to display the interiorof a vessel within one image. To accomplish this, a re-samplingstrategy different from existing CPR methods is introduced. CPRtechniques display re-sampled data in close vicinity of a surfaceintersecting the volume. This surface is defined by the central-axisof the vessel and avector-of-interest(vi) (see Figure 2a). The lattermight be defined in a local coordinate system of the central-axis.In any case the data are re-sampled in a linear way defined by thevectorvi. The vector-of-interest describes the re-sampling directionwhich is orthogonal to the viewing direction.











    (a) (b) (c)

    Figure 3: Spiral-of-interest (a), constant angle sampling (b), con-stant arc-length sampling (c)

    The helical CPR method is based on a non-linear re-sampling ofthe data. The vector-of-interest as generating element for the sur-face is replaced by aspiral-of-interest(si) (see Figure 2b). Thisresults in a convoluted surface around the central-axis. At a suf-ficiently small distance between each winding the vessel is inter-sected several times. Stenosis, calcification, and occlusions are in-cluded in the computed surface. This surface is helical and dis-played.

    3.1 Method DescriptionThe helical CPR visualization technique is based on the straight-ened CPR method. Along the central-axis of the vessel cross-sections are calculated at an appropriate sampling distance. Withineach section a local 2D coordinate system is defined. The centerof the cross-section represents the estimated center of the vessellumen from the corresponding center-line position. Starting fromthis center point two interleaved spiralss1 ands2 are computed (seeFigure 3a). In order to maintain a uniform sampling of the ves-sel cross-sections a spiral with constant inter-winding distance wasselected. This requirement is satisfied by the Archimedean spiralwhich can be notated in polar equation as:

    r = aθ (1)

    The transformation of points on the curve into Cartesian coordi-nates is straightforward. Thus the computation of each pointXs1 ons1 andXs2 ons2 is performed as follows:

    Xs1 = aθ(


    )Xs2 = aθ

    (cosθ ′

    sinθ ′

    )whereθ ′ = θ +π

    (2)For an appropriate sampling of the vessel lumen the parameterawas set to1/π. This assures the constant distance between thewindings of the two interleaved spirals to be one. The computedpointsXs1 andXs2 on the spiral are transformed back into volumespace and re-sampled.

    The center of each row in the final CPR image corresponds tothe center of the vessel cross-section. Starting from this referencepoint the image space to the left is filled with data re-sampled bys1and to the right with data froms2.

    3.2 Sampling StrategyThe current implementation of the helical CPR technique supportstwo sampling strategies for computing points from the spiral. Theangleθ is increased by a constant angleω for each point in the caseof constant anglesampling (see Figure 3b). For each sampling stepa constant distance∆ on the arc-length of the spiral is covered ifconstant arc-lengthsampling is applied (see Figure 3c).

  • Constant angle. For each point re-sampled from the spiral thegenerating angleθ is incremented by a fixed angleω. Each windingis rendered into an equal sized area in the final image. Therefore thecomparably densesampled area in close vicinity of the vessel centeris amplified in image space. The resulting fish-eye zooming effectis achieved at the cost of increased distortion.

    Constant arc-length. Given a fixed sampling distance∆ be-tween two adjacent points on the spiral the incrementω of the angleθ is approximated. The incrementω is defined by the ratio of∆ andthe circumference calculated from the most recent radius. As usu-ally small sampling distances are used the error introduced by thisapproximation is negligible. The extent of the vessel in the CPRimage is directly proportional to the area of the vessel lumen. Thuslarge vessels occupy a superproportional large image space.

    3.3 Phantom DatasetThe images generated by helical CPR visualization provide a quiteunconventional display of vascular structures. Therefore a phantomdataset containing a set of typical vascular abnormalities was com-puted. By means of this dataset the typical pattern of each anatom-ical case are demonstrated (see Figure 4).

    The simulated vessel of the phantom dataset contains a semi-circumferential vessel wall calcification, and a circumferetial ves-sel wall calcification without luminal stenosis. An eccentric ”soft-plaque” stenosis and a segmental concentric high-grade stenosiswas also simulated, respectively (top-down order). Figure 4a showsa direct volume rendering display of this simulated vessel. Fromthis visualization method the flow channel within the circular ves-sel wall calcification can not be gauged. A coronal and sagittalstraightened CPR is presented in Figure 4b and c. The partial ves-sel wall calcification as well as the eccentric stenosis are not visiblein the sagittal straightened CPR. This demonstrates the need for dif-ferent longitudinal sections. Figure 4d and e show the result of thehelical CPR technique with constant arc-length and constant anglesampling. The vertical black lines indicate the image space requiredfor one winding. Partial vessel abnormalities tend to reoccur in aseparated fashion several times in the image whenever the area isintersected by a winding. Vessel pathologies clearly stand out inthe helical CPR images. The helical CPR ”peels-off” the bloodves-sel. Thus the extent of abnormalities is enhanced.

    3.4 ResultsThe application of an helical CPR technique on a real world datasetis presented in Figure 5. A constant angle and constant arc-lengthsampled helical CPR display is compared to a straightened CPR im-age. In contrast to the straightened CPR the helical CPR shows thearea of vessel lumen instead of the diameter. The white arrow illus-trates an example where the helical CPR outperforms a traditionalCPR. The small flow-channel of the stenosis is not touched by thedisplayed longitudinal section of the straightened CPR and there-fore not visible. However in both helical images this flow-channelis displayed. As eccentric lesions cause repetitive patterns in theimage space, the attention of the observer is immediately drawn tothose areas even if a lesion is not visible in a standard CPR display.


    The aim of untangled CPR visualization is to display a vascular treewithout overlapping arteries independently from the spatial posi-tion of the intersecting plane. In order to accomplish this require-ment the spatial relations of the projected vessels have to be re-laxed. Branching points of the vessel tree are taken as pivot points.

    Rotating the corresponding vessels around these pivot points in im-age space eliminates vessel overlaps. Keeping the introduced dis-tortions small maintains fast perception and reduces the impact ofre-sampling artifacts in the final image. Thus the applied transfor-mations are restricted to the branching points (bifurcations) of thearterial tree. In addition to that this transformations should be ap-propriate in terms of changing the tree layout and appearance with-out violating the non-intersection criterion. Thenon-intersectioncriterion is defined in a way that two vessel hulls must not crosseach other at any time. The definition of a vessel hull is describedbelow.

    The new CPR technique is based on the stretched CPR and multi-path CPR method introduced by Kanitsar et. al. [6]. This methodmaintains the curvature of the vessel to a high extent and preservesisometry. The vessel’s curvature is crucial in approximating a ge-ometrically undistorted appearance of the vessel tree. Isometry onthe other hand is important for vascular investigations. Stent plan-ning and stenosis grading estimation are possible in the final image.

    The input of the algorithm is a tree graph representing the topol-ogy of the vascular structure. For each vessel segment the centerline of the vessel is stored as a set of adjacent points at an appropri-ate sampling distance. In practice it turned out that diameter estima-tions of vessels are not reliable enough in certain cases. Thereforefor the purpose of generality the algorithm does not take diameterinformation into account. However, the adaptation to this additionalinformation would be straightforward.

    4.1 Method outlineThe untangling CPR method consists of four main steps. As alluntangling calculations are performed in image space the tree graphis first mapped to image space using a stretched CPR projection. Ina consecutive step the transformation of each subtree with respect tothe non-intersection criterion is performed. Afterwards the imagespace is partitioned in a way that each vessel obtains those parts ofthe image space which are closest in scan-line direction. Finally theimage is rendered.

    Tree projection. The vascular tree is mapped to a projectionplane coplanar to the viewing plane. For each two successive pointsthe subsequent point is rotated around an axis defined by the previ-ous point and the vector-of-interest. This rotation is carried out foreach point starting from the root of the vascular tree. This producesa stretched vascular tree.

    Untangling operation. From the projected tree graph in imagespace the necessary transformations for each node are calculated.This is done by recursively circumscribing the subtrees with vesselhulls. The first pass is bottom up maintaining only the correct trans-formation of the largest enclosing vessel hulls. In a second pass thefinal transformation for each vessel hull segment is accumulated topdown.

    Image space partitioning. Before rendering the final imagethe extent of each vessel segment is cropped in a way that no over-lapping image areas remain. This process determines the startingpoint and the end point of each scan line for rendering.

    Rendering. Each vessel strip is rendered separately. Concep-tually the vessel strip is first extracted from the dataset using astretched CPR mapping. Afterwards the strip is transformed to theposition defined by the untangling process. In a further step the stripis clipped according to the space partitioning information. Finallyeach cropped scan line is rendered into the image.

  • Figure 4: Phantom dataset: Direct volume rendering (a); Coronal (b) and sagittal (c) straightened CPR; Helical CPR with constant arc-length (d) and fixed angle (e).

    Figure 5: A straightened CPR (b), an helical CPR with constant angle sampling (a), and constant arc-length sampling (c) of a real worlddataset.

  • P0




    = Hcenter

    H right’ H left’

    Hright Hleft

    vright v right’vleftv left’

    e e

    Figure 6: The vessel hull primitive

    4.2 The vessel hull primitiveThevessel hullis the basic primitive for further intersection tests. Itencloses the vessel’s centerline in image space as shown in Figure 6.The centerline is given as a set of pointsP = {P0..Pn−1}. A vesselhull is a sector of a circle. The root of a vascular subtree defines thecenter pointHcenter. A matrix ℜHcenter associated with each centerpoint describes a rotational transformation of the subsequent tree.The pointsH ′right andH

    ′le f t result from a conservative estimation

    of the leftmost and rightmost extent of the enclosing subtree seenfrom the center point.

    The vessel hull encloses the centerline of the vessel thus twoneighboring vessels touching each other is not prevented by thisprimitive. To overcome this situation the vessel hull is enlarged bya small angleε as depicted in Figure 6. Depending on the size ofthe inspected vessels thisε may be adjusted by the user on the fly.However anε ≤ 2◦ was found to be appropriate for most testeddatasets. If the vessel diameter is known at the extremal pointsPiandPj thenε can be easily calculated more accurately. The pointscomprising theε-tolerance are referred to asHright andHle f t.

    Each vessel hull primitive can be described as a tuple ofV = {Hcenter, Hle f t, Hright , ℜHcenter} where~vle f t =

    −−−−−−−→HcenterHle f tand~vright =

    −−−−−−−−→HcenterHright .

    4.3 Putting things togetherA rule based approach is applied for combining vessel hulls fromvarious parts of the vascular tree (see Figure 7). The projected ves-sel tree is approximated by an enclosing hierarchy of vessel hulls.This hierarchy is constructed bottom up. Combining two vesselhulls at a branching point involves a rotation around the commonpivot pointHcenter. This results in a new larger hull which boundsthe entire subtree. Constructing the vessel hull hierarchy involvesseveral cases wich are discussed in the following. A vessel hullcreated from the vessel centerline is based oncase 1. Neighbor-ing vessel hulls are combined according tocase 2. An enclosingvessel hull from two consecutive vessel hulls is created incase 3.The assembling process is done bottom up. This results in a binarytree of vessel hulls where each node is represented by a vessel hullcircumscribing all subjacent vessel hulls.

    Case 1. The center pointHcenteris defined by the first pointP0 ofthe vessel segment. The first point on the convex hull of the vesselsegment in clockwise orientation is denominated as pointPi and incounterclockwise direction as pointPj . Because of the stretchedCPR mapping from volume to image space, pointPn−1 is the pointwith maximum distance fromP0. Thus the radius of the vessel hullis computed as|Pn−1−P0|. The normalized vectors~vle f t and~vrightrepresent the directions fromHcentertoPi andPj respectively. Thesevectors are scaled according to the radius of the vessel hull. The


    center= H2



    right H1













    right H1








    Case I Case II Case III

    Figure 7: The three vessel hull combination cases.

    tolerance angleε is incorporated by a transformation withℜε andℜ−ε . Finally Hright andHle f t are computed.

    Case 2. For the case of two adjacent vessel hull primitives (V1andV2) (see Figure 7) the ordering of the subtrees has to be de-termined first. According to this decision the left vessel hull withrespect toHcenter= H1center= H

    2center is denominated asV

    l and theright one asVr . If the vessel hull primitives overlap, an untanglingangleγ is computed from~vrle f t and~v

    lright (see Figure 11a). From

    this angle the rotational matricesℜHrcenter andℜH lcenter are calculated.These matrices define a transformation of the vessel hull primitivesVr andV l in a way that the primitives do not overlap anymore. Thisimplies a transformation of the associated vascular subtree (see Fig-ure 11a). If the vessel hull primitives do not overlap the matricesare simply the identity matrix.

    The vectors~vle f t and~vright of the combined vessel hull are com-

    puted from the transformed vectors~vlle f t and~vrright . The radius of

    the enclosing vessel hull is defined by the maximum radius ofV1

    andV2. From this informationHle f t andHright is computed. Thenewly generated vessel hull encloses the non-overlapping underly-ing vessel hulls.

    Case 3. The combination of two successive vessel hulls isstraightforward.V1 is considered to be the predecessor ofV2 asdepicted in Figure 7.H1center is taken as the new center pointHcenterof the enclosing vessel hull. The direction to the rightmost point ofH1right andH

    2right with respect toHcenter is considered to be~vright .

    Vector~vle f t is calculated likewise. The radius of the new vessel hull

    V is defined by the maximum distance fromHcenter to H1right , H

    1le f t,

    H2right , andH2le f t.

    Case 3 is similar to the2nd case. However no additionaltransformation is introduced during this step. This is becauseundersampled areas occur under this circumstances as shown inFigure 11b.

    Any treelike vascular structure can be processed using thisset of rules. One bifurcation for instance is subdivided into threecase 1, one case 2and, onecase 3. The recursive algorithmfinishes with a hierarchy of enclosing vessel hulls where the roothull contains the entire vessel tree. A detailed description of eachassembling case in abstract notation is given in Figure 8.

  • Case 1:Hcenter← P0~vright ← Pj −Hcenter,{Pj |Pj , ` ∈ P∧∀`(` left of −−−−−→HcenterPj )}~vle f t ← Pi −Hcenter,{Pi |Pi , ` ∈ P∧∀`(` right of −−−−−→HcenterPi)}rad ← |Pn−1−P0|Hright ← Hcenter+ rad · (ℜε ?~vright/|~vright |)Hle f t ← Hcenter+ rad · (ℜ−ε ?~vle f t/|~vle f t|)

    Case 2:Hcenter ← H1center(V l ,Vr )← if (changeOrder)then (V2,V1) else(V1,V2)ℜHrcenter← RotationMatrix(Hcenter, –0.5max(^(~v

    rle f t,~v

    lright ),0))


    ← RotationMatrix(Hcenter, +0.5max(^(~vrle f t,~vlright ),0))~vright ←~vrright~vle f t ←~vlle f trad ← max(radl , radr )Hright ← Hcenter+ rad · (ℜHright ?~vright/|~vright |)Hle f t ← Hcenter+ rad · (ℜHle f t ?~vle f t/|~vle f t|)

    Case 3:Hcenter← H1center~vright ← if (H2right left of


    1right ) then H

    1right elseH


    ~vle f t ← if (H2le f t right of−−−−−−−→HcenterH

    1le f t) then H

    1le f t elseH

    2le f t

    rad ← max(rad1, |Hcenter−H2right |, |Hcenter−H2le f t| )Hright ← Hcenter+ rad ·~vright/|~vright |Hle f t ← Hcenter+ rad ·~vle f t/|~vle f t|

    Figure 8: Assembling of vessel hulls.

    4.4 Layout definitionThe decision of the vessel hull ordering incase 2has a significantimpact on the layout of the displayed vascular tree. Two differentapproaches have been investigated (see Figure 9). One possibil-ity is a left-to-right ordering based on the spatial relations of thevascular tree according to the currently used viewing direction. Inthis case the ordering is based on the spatial location of the vesselhulls according to the current viewing direction. This approach isreferred to asadaptive layout. In contrast to that afixed layoutwasinvestigated too. The fixed layout is independent from the viewingdirection. In the present prototype implementation the vessel order-ing of the coronal display is maintained for all viewing directions.For clinical routine applications a standardized ordering might bea reasonable solution. According to this standardization the left-most artery for peripheral CTA examinations might be the left deepfemoral artery (profunda femoris) followed by the left anterior tib-ial artery and so on.

    The advantage of an adaptive tree layout is a more efficient uti-lization of image space and less distortion introduced by the untan-gling operation compared to a fixed layout. The introduced distor-tion measured by the sum of untangling anglesγ over a 360◦ rota-tion of a peripheral CTA dataset is depicted in Figure 10. However,in the case of an adaptive tree layout discontinuities occur when-ever the ordering of the vessels change. Thus the inspection of asequence of images becomes more difficult as the identification ofeach vascular segment becomes more difficult. Labelling the vesselsegments eases this limitation for examinations carried out on a setof images.

    4.5 Image space partitioningBevor rendering the final image the image space has to be separatedinto fragments for each vascular structure. This step is necessarybecause arteries are possibly obscured by bones if multiple layersare rendered into a single image area (see [6] for details).

    A principle drawing direction is associated with each vessel seg-ment. Each point of the projected vessel center line maintains ascan linededuced from this principal drawing direction. These scan



    viewing direction

    fixed la


    adaptive layout









    Figure 9: Different layout definition








    0 45 90 135 180 225 270 315


    AdaptiveΣ γ

    Angle of viewing direction


    Figure 10: Comparison of introduced distortion.

    lines are transformed according to the rotational matrixℜHcenter (seeFigure 11a). The scan lines represent those parts of the image inwhich the corresponding re-sampled data from the dataset are ren-dered into. Because of the applied transformations the scan linesneed not be aligned to the pixel rows of the image.

    In order to avoid overlapping areas an appropriate start and end-point for each scan line has to be determined. For this reason aspecial sort of distance map is build up every time an image is ren-dered. This distance map is defined by the projected center linesof the vessel tree and its scan lines. In contrast to the traditionaldistance map the distance metric is not defined by an Euclideandistance but by a distance along scan lines. The result of this op-eration is a fragmented image space where each vessel segment isassigned a maximally image area.


    (a) (b)

    undersampled area

    scan lines

    Figure 11: Image space partitioning

  • 4.6 Results

    A comparison of stretched multi-path CPR and untangled CPR ispresented in Figure 12. In an anterior view the results of the com-pared methods are rather similar, because there are only few over-lapping arteries. In the case of a lateral view the multi-path CPRdisplay provides hardly any diagnostically relevant information.Many superimposed arteries obscure each other. In comparison tothat the untangled CPR still provides an unobscured view of the en-tire vascular tree. Each vessel segment is displayed in diagnosticquality. For instance the stenosis, heavy calcification, and a longocclusion in the left superficial femoral artery is clearly visible. Forboth untangled CPRs a fixed layout was used.

    Figure 13 presents a sequence of untangled CPR images from anabdominal CT-angiographic dataset. A fixed layout was used, and a1D transferfunction was applied to the re-sampled data approximat-ing the tissue color of the anatomical structures. Areas of arterialwall calcification are shown in the distal abdominal aorta (225◦)and in the left internal iliac artery (270◦). The abdominal aorta isthe main vessel in the upper half of the image. It is branching intothe left and right iliac (pelvic) arteries where left and right in the im-age is interchanged due to the patient orientation. The pelvic arterybranches into the thicker external pelvic artery and the thinner innerpelvic artery.

    Even though the examination is intended to be done on a smallset of pre-computed images, the performance of the algorithm isacceptable for applications in the clinical workflow. The displayeduntangled CPR image in Figure 12 was calculated from a real worlddataset with a scanned resolution of 512x512x988 voxel and anoriginal image size of 1164x1097 pixel. The average rendering timeper image of the current Java based implementation on a PC work-station with an Intel PIII 1GHz main processor took 2.3 seconds.


    Two methods for efficient vessel visualization were proposed. Inboth methods the spatial coherence of the investigated objects wasrelaxed to a certain degree.

    The helical CPR technique is a new way to display the volume ofvascular structures. The motivation for this visualization techniquewas not an imitation of the natural appearance of the object but thedisclosure of diseased vessel segments. Helical CPR visualizes theinterior of a vessel in a single image. Re-sampling a spiral aroundthe vessel central axis does not involve any compositing of sam-ples. Therefore no features of the vessel may be hidden by otherstructures as long as the vessel area is sampled densly enough. Thisexplorative study has shown that the detection of vessel abnormali-ties is possible from such visualizations which literally peel-off thevessel volume. A particularly attractive future application of thistechnique is visualization of coronary artery disease from recentlydeveloped electrocardiographically gated CT.

    The untangled CPR has significant advantages over existingmulti-path CPR techniques. This new technique produces an un-obscured display of a vascular tree, independent of the viewing di-rection. Small rotations around the branching points of a vessel treeeliminate occlusions. Therefore the size of the introduced distortionis kept small.

    Even though the use of image space is not optimal, the mainrequirement of unobscured display of vessels from any viewing di-rection is fulfilled. In addition, untangled CPR preserves isometrywhich is an important requirement for vascular lesion assessment.The potential for clinical application of this technique is obviously– a more efficient way to assess any complex arterial trees for thepresence and extent of vascular disease. Clinical validity and appli-cability to other vascular territories are currently investigated.


    The authors would like to thank Michael Knapp for the generationof the phantom dataset. The work presented in this publication hasbeen funded by theADAPT project (FFF-804544).ADAPT is sup-ported byTiani Medgraph, Vienna (, and theForschungsf̈orderungsfonds für die gewerbliche Wirtschaft, Aus-tria. See for furtherinformation on this project.


    [1] S. Achenbach, W. Moshage, D. Ropers, and K. Bachmann.Curved Multiplanar Reconstructions for the Evaluation ofContrast-Enhanced Electron-Beam CT of the Coronary Ar-teries. InAm. J. Roentgenol., pages 895–899, 1998.

    [2] Brian Avants and James Williams. An Adaptive Minimal PathGeneration Technique for Vessel Tracking in CTA/CE-MRAVolume Images. InMICCAI 2001, pages 707–716, 2000.

    [3] P. Felkel, A. Fuhrmann, A. Kanitsar, and R. Wegenkittl.Surface Reconstruction Of The Branching Vessels For Aug-mented Reality Aided Surgery. InBIOSIGNAL 2002, pages252–254, June 2002.

    [4] Horst Hahn, Bernhard Preim, Dirk Selle, and Heinz-Otto Peit-gen. Visualization and Interaction Techniques for the Explo-ration of Vascular Structures. InIEEE Visualization 2001,pages 395–402. ACM, October 2001.

    [5] Sha He, Ruping Dai, Bin Lu, Cheng Cao, Hua Bai, and Bao-lian Jing. Medial Axis Reformation: A New VisualizationMethod for CT Angiography.Academic Radiology, 8:726–733, 2001.

    [6] Armin Kanitsar, Dominik Fleischmann, Rainer Wegenkittl,Petr Felkel, and Meister Eduard Gröller. CPR - Curved Pla-nar Reformation. InIEEE Visualization 2002, pages 37–44.ACM, October 2002.

    [7] Armin Kanitsar, Rainer Wegenkittl, Petr Felkel, DominikFleischmann, Dominique Sandner, and Eduard Gröller. Com-puted Tomography Angiography: A Case Study of PeripheralVessel Investigation. InIEEE Visualization 2001, pages 477–480. ACM, October 2001.

    [8] A. Koechl, A. Kanitsar, F. Lomoschitz, E. Groeller, andD. Fleischmann. Comprehensive assessment of peripheralarteries using multi-path curved planar reformation of CTAdatasts. InEurop. Rad., volume 13, pages 268–269, 2003.

    [9] Geoffrey Rubin, Andrew Schmidt, Laura Logan, and MarkSofilos. Multi-Detector Row CT Angiography of Lower Ex-tremity Arterial Inflow and Runoff: Initial Experience. InRa-diology 2001, pages 146–158, 2001.

    [10] O. Wink, W. Niessen, and M. Viergever. Fast Delineation andVisualization of Vessels in 3-d Angiographic Images.IEEETransactions on Medical Imaging, 19:337–346, 2000.

    [11] C. Zahlten, H. Juergens, and H. Peitgen. Reconstructionof Branching Blood Vessels from CT-Data. InEurograph-ics Workshop on Visualization in Scientific Computing, pages161–168, 1994.

  • (a) (b) (c) (d)

    Figure 12: A peripheral CTA dataset rendered from coronal and sagittal view using stretched multi-path CPR (a, c) and untangled CPR (b, d)respectively (fixed layout)

    Figure 13: A color coded sequence of untangled CPR images from different re-sampling directions.