Hierarchical Volume Visualization of Brain Anatomy

Jean-Paul Balabanian1, Martin Ystad2, Ivan Viola1, Arvid Lundervold2, Helwig Hauser1,and Eduard Groller3,1

Department of{Informatics1 | Biomedicine2}, University of Bergen, NorwayVienna University of Technology, Austria3

Email:{firstname.lastname}@{biomed|ii}.uib.no, groeller@cg.tuwien.ac.at

Abstract

Scientific data-sets often come with an inherent hi-erarchical structure such as functional substructureswithin organs. In this work we propose a newvisualization approach for volume data which isaugmented by the explicit representation of hier-archically structured data. The volumetric struc-tures are organized in an interactive hierarchy view.Seamless zooming between data visualization, withvolume rendering, and map viewing, for orienta-tion and navigation within the hierarchy, facilitatesdeeper insight on multiple levels. The map showsall structures, organized in multiple hierarchy lev-els. Focusing on a selected node allows a visualanalysis of a substructure as well as identifying itslocation in the hierarchy. The visual style of thenode in focus, its parent and child nodes are auto-matically adapted during interaction to emphasizethe embedding in the hierarchy. The hierarchy viewis linked to a traditional tree view. The value of thisnew visualization approach is demonstrated on seg-mented MRI brain data consisting of hundreds ofcortical and sub-cortical structures.

1 Introduction

Research in information visualization has many ex-amples of visualizing hierarchical data such as treesand graphs. Scientific data often has an inherent hi-erarchy that is in many cases not fully exploited dur-ing visualization. In the medical domain it is ofteneasy to describe the inherent hierarchical nature ofthe data. The human body can be semantically di-vided into several structures that have a hierarchicalrelationship with each other. For example the armcan be substructured into upper arm, forearm, andhand. The hand can be further divided into fingersand palm. Another example of a hierarchical struc-

ture, and also the one we are here focusing on, isthe brain. The anatomical hierarchical subdivisionof the brain starts with the separation of the left andright hemispheres, then the cortical and sub-corticalareas, followed by subdivision into different lobes,consisting of several gyri and other structures.

In medical education it is difficult to convey this3D spatial relationship by the use of textbooks.Thus, medical students have to perform training oncadavers in order to acquire this kind of knowledge.The amount of information that is possible to ex-tract from a textbook is to a significant amount re-lated to the contained illustrations. The amount ofknowledge gained from cutting into a real brain isalso limited. Cutting open one structure to study itssub-structures will make the higher level structureunusable for further studies due to its irreversiblemodification. It is also possible to study brain databy looking at MRI slices, but analyzing such slicesrequires reasonable expertise. 3D volume visualiza-tion can help in visualizing the structures. Howeverit is difficult to infer hierarchical and semantic infor-mation from these visualizations, especially whenmany structures are to be investigated.

Our approach is based on two different types ofdata. One is 3D anatomical data from MRI, with bi-nary segmentation masks, and the other is abstracthierarchical information inferred from the 3D data.The proposed approach in this paper tries to notonly show the anatomical structure but to integratehierarchical semantics and volume information inthe same visualization. The visualization combinesthe field of scientific visualization with informationvisualization by rendering a hierarchical layout inthe same view as the volume rendering. Figure 1shows a closeup example of this combined view.

The major contribution of this work is the com-bined visualization of scientific volume data withinherent hierarchies. We provide a seamless inter-

VMV 2008 O. Deussen, D. Keim, D. Saupe (Editors)

face that enables an integrated interaction betweenabstract hierarchies and scientific data. We do thisby creating an overview map where the hierarchyof the data is represented and where it is possibleto zoom in to reveal knowledge about the volumedata. At the volume data level we change the visualrepresentation of structures with auto-styling so thatthe hierarchical relationship between the structuresbecomes evident in the spatial domain. Using thenovel concept of raycasting portals we are able torender more than 150 structures with volume ren-dering at the same time.

This paper is organized as follows: In the fol-lowing section we present an overview of existingvisualization techniques that relate to our work. InSection 3 we describe our approach to visualize hi-erarchical data and present results in Section 4. InSection 5 we discuss the results and mention futurework. Finally we conclude in Section 6.

2 Related Work

We aim at volume rendering with the look-and-feelof medical textbooks such as the anatomical atlasby Sobotta [16]. Volume rendering has become alarge field of research. The GPU-based renderingapproaches that we build on are described by En-gel et al. [3] and Kruger and Westermann [8]. Theillustrative results are produced with style transferfunctions as proposed by Bruckner and Groller [2].

Hierarchical data are easier to navigate and togain knowledge from if an appropriate interactionmetaphor and visualization is used. The evaluationdone by Wang et al. [19] confirms this. Hierarchicalinformation is often visualized as a tree. The infor-mation visualization community has done extensiveresearch in the field of visualizing and navigatinghierarchical data. One type of approach maximizesthe utilization of the available screen-space, calledspace-filling techniques, such as tree-maps [14, 15],information slices [1], and the InterRing [21]. Theyalso indicate size measures associated with the data.In visualizing a file-system, for example, the tree-map technique uses the size of a file or directoryas a measure of the size of a structure. Since thedata is hierarchical, the size coding is applied recur-sively and the space occupied by a parent node issubdivided by its children. The visualization used toshow the parent-child relationship is depicted withrectangles inside rectangles. Similarly, information

(a)

(b)

Figure 1: (a) The occipital lobe is colored to indi-cate the hierarchical relationship usingauto-styling.(b) Interactive change of the visual representation ofone structure in the cingulate cortex.

slices and InterRing use cascading circles and visu-alize the size measures as sector pieces.

Other techniques visualize trees without givingan indication of the relative sizes of the hierarchies.Cone Trees [13] and hyperbolic trees [10], for ex-ample, create a navigatable space with nodes in 3D.Other techniques such as RINGS [18] and Balloontrees [9] position nodes radially in 2D. The latter ap-proaches have some similarities with our techniqueto lay out hierarchical data.

Interaction with and navigation of hierarchical

(a) (b) (c)

Figure 2: Hierarchically boosted volume visualization. (a) The data basis.(b) Components of the hierar-chical visualization. (c) Interaction metaphors.

data is also a topic of research. An example is auto-matic panning and zooming [20] which efficientlymoves from one node to another while preservingthe overview by immediately zooming out to showthe context. Other approaches let the user focus onsome region of interest. InterRing lets the user ex-pand a hierarchical level of interest. The other lev-els are reduced automatically to accommodate theregion of interest. Another approach to visualizingthe region of interest especially, was suggested byStasko and Zhang [17]. The outer or inner part of aradial visualization is used as a special area to ren-der the region of interest. Another interesting wayof performing focus+context visualization is basedon non-linear magnification lenses [7].

Other visualization techniques try to make thetree and graph visualizations more sparse by re-ducing the number of connection lines betweennodes. Examples are edge bundling [6], or chang-ing the thickness of connection lines such as in arc-trees [11]. Herman et al. [5] provide an exhaustivesurvey on trees and hierarchy interaction.

FreeSurfer [4] is a set of tools for the study ofcortical and sub-cortical anatomy. It provides auto-mated parcellation of the cerebral cortex and label-ing of sub-cortical tissue classes in MRI volumes.

Previous work that proposes techniques to visu-alize the hierarchical nature of the brain has beenproposed by Pommert et al. [12]. Their techniqueconsiders several different types of hierarchies. Theuser has to actively select a structure, then selectwhat type of hierarchical information is interestingfrom a popup menu. Our technique differs signifi-cantly from their approach. In our visualization, for

example, the hierarchy is the context that the useris navigating in. When focusing on a feature morehierarchical information is automatically provided.

Sources that describe techniques to combine hi-erarchy visualization and scientific visualization inthe same context are scarce. The closest solutionto resemble our technique is volume rendering ofsegmented structures with one structure highlightedand the other structures as context.

3 Spatial Data with Hierarchical Se-mantics

We integrate two spaces, an abstract space with ahierarchy and a data space where the volume data isdefined. We enable seamless zooming between thehierarchical model and the anatomical data in thespirit of the focus+context metaphor. We propose atree layout of the hierarchical data where each nodeshows a volume rendering of the semantically asso-ciated structure and a descriptive label. We call thisthecontext view. It is crucial that rendering and nav-igation of this view is interactive. The navigationincludes zooming from the context view to the vol-ume data and the hierarchically guided explorationof this view. Figure 2 illustrates the different aspectsof our approach. Figure 2(a) shows the availabledata basis, in Figure 2(c) the interaction possibili-ties are shown, and in Figure 2(b) the visualizationtechniques that create the final results are shown.The visualization changes according to the user in-teraction. Some of the visualization techniques areonly active during specific interactions, others are

(a) (b)

Figure 3: Layout patterns for the context view. (a) Afan pattern, filling the area of a sector. (b) A clusterlayout around a replicated group node in the centerwith a dashed line connecting it with its original.

active during the entire interaction process.

3.1 Hierarchy in the Data

In Figure 2(a) three types of data are listed as in-put for our approach. The data which we visualizedhere is theBert data-set as provided by FreeSurfer.The data is T1 weighted MRI and FreeSurfer auto-matically generates binary segmentation masks formany structures in the brain. This process is basedon an brain atlas technique and it takes approxi-mately 20 hours. The segmentation masks that weuse are the ones generated for the cortex and thesub-cortical areas of the brain (APARC+ASEG).The segmentation masks represent small regionsand structures that by themselves do not form a hi-erarchy. We have created a hierarchical tree thatassociates segmentation masks with labels and la-bels with groups that are semantically meaningful.For example cortical ridges, denoted as gyri, aregrouped together to form larger structures calledlobes. These groups are part of other groups, suchas lobes that are part of the cerebral cortex. The hi-erarchical groupings stop with the hemispheres ofthe brain. The resulting data-set contains a hier-archical overview that anatomically and hierarchi-cally describes the brain and its structures.

3.2 Hierarchy Visualization

The hierarchical information of the data is visual-ized in two ways. First, a context view is generatedthat illustrates the hierarchical structure. A node-link diagram is used. Every node has links to its

children and to its parent and all nodes are labeled.The automatic layout scheme attempts to create acontext view that takes advantage of symmetries tosupport the orientation of the user. Second, childand parent nodes are displayed close to each other.A unique coloring of sub-structures indicates the hi-erarchical structure. An example of this can be seenin Figure 1(a).

We have defined semantics for the hierarchicalinformation associated with the data. A leaf node isa segment node. These nodes have a direct corre-spondence with a segmentation mask. Nodes withchildren are group nodes, and group nodes that onlycontain leaf nodes are called leaf-clusters. Thissemantic is used for context view creation, auto-styling, and volume rendering. The leaf nodes ofa leaf-cluster are shown in Figure 3(b). In the samefigure the two nodes connected with a dashed lineare group nodes.

To optimize screen-space utilization we placestructures as close to each other as possible, whenwe create the context view. In addition, structuresshould be positioned in such a way that the hier-archical relations are self evident. The user shouldfast and easily recognize the different structural fea-tures as generated in the context view.

The design choices of the context view are de-rived from the hierarchical nature of the data (Fig-ure 4(a)). We place structures in a radial pattern andassign sub-structures to fractions of the sectors thatare occupied by higher level structures. The left halfand the right half of the layout correspond to theleft and right hemispheres of the brain. Each quad-rant represents a high level feature. These featuresare the cortex in the upper quadrants and the sub-cortical areas in the lower quadrants. Sub-structuresof these features are given as fractions of these foursectors. Groups get a fraction of their parent’s sec-tor based on the number of children.

Some of the structures at the leaf level in the hi-erarchy have many siblings. In the case of leaf-clusters we want to minimize the occupied screen-space. Nodes in a leaf-cluster are positioned arounda central point without any overlap. For helping theuser to keep the context in mind, the group nodeof the leaf-cluster is replicated in the center of thecluster (Figure 3(b)). The clustering reduces thearea occupied by sibling nodes relative to the sec-tor size. If we do not cluster the sibling nodes inthis manner the space needed to draw all the sibling

(a)

(b)

Figure 4: A context view (a) showing the entire hi-erarchical layout with a close up below (b). Theright hemisphere is indicated in (a) as the dashedrectangle. Some features have been labeled in thefigure to indicate their position. The close up of theregion enclosed by the orange dashed rectangle isshown in Figure 1.

nodes increases, causing the nodes to move furtheraway from the original parent node. Groups thatonly have group nodes or a combination of groupnodes and leaf nodes as children are placed in a fanpattern, positioned at a distance where nodes do notoverlap. This means that the nodes must be movedto a distance where all nodes can be positioned onan arc within the sector bounds. The group nodeis in this case replicated as well but it is positionedbetween the fan and the original group node (Fig-ure 3(a)). At the two highest levels. i.e., the brainand the two hemispheres, we do not replicate thegroup nodes. The size of the rendered replicatedgroup node is adjusted depending on the number ofsiblings. In a leaf-cluster, for example, the repli-cated group node increases in size when the numberof siblings creates a circle that is much larger thanthe minimum node size (Figure 8).

The position where we place a replicated groupnode is also the position we use to bundle the con-nection lines between parent and children. In Fig-ure 3 this can be seen as the connection lines fromleaf nodes and group-nodes to the replicated groupnode of their parent. This makes the overview lesscluttered. Drawing one line between a group nodeand a replicated group node and then one line fromeach child to the replicated group node is morespace efficient than one line per child node to theoriginal group node.

The Figure 4(a) shows the complete hierarchyof the brain with every node rendered in a ring.The ring is rendered as two concentric circles withdifferent radii. The orange dashed square in Fig-ure 4(b) represents the zoomed area given in Fig-ure 1. In this closeup we can see most of the dif-ferent types of visualizations from our approach. InFigure 1 there are two structures organized in twoleaf-clusters, i.e., the occipital lobe and the cingu-late cortex. When a node is highlighted, its ringis rendered in blue gradients. In case of replicatedgroup nodes the highlighting is done for both nodes.Each node has an associated label which is placedon a curve on the top half of the ring.

When the user is looking at the context view,lines provide hierarchical information, such aschild-parent relationships. The algorithm drawslines recursively between parent nodes and childnodes. The line is drawn from the border of thering. In the case of replicated nodes, dashed linesare drawn between the original node and the repli-

cated node. This is shown in Figure 1 and Figure 4.An intuitive way of interaction with the context

view is the possibility to focus on nodes. There aretwo types of direct interaction which result in focus-ing, i.e., hovering over a node and selection. Whenthe mouse hovers over a node the node will initiateauto-styling, described in Section 3.4. Selection isdone through clicking on a node. This centers theview on the node and enables manual selection ofthe style of that structure (Section 3.5).

3.3 Raycasting Portals

In the center of every node a volume renderingof the associated structure from the hierarchy isshown. This depiction of the volume data is gener-ated by a GPU-based volume raycasting technique[8]. A proxy geometry is used to render the cubi-cal shape of the volume data. We use the 3D tex-ture coordinates to render colors that we use as amap into the volume data. The structures to visu-alize usually occupy only a sub part of the volume.If we only render the voxels included in the seg-mentation mask, we can significantly improve ren-dering time. We calculate the bounding boxes forall segmentation masks and the bounding boxes forall nodes higher in the hierarchy and use this to re-shape the proxy geometry. We offset the texture co-ordinates so that they map to the coordinates of thebounding box of the structure. The aspect ratios ofthe proxy geometry are adjusted to match the as-pect ratios of the bounding box. An example of thereshaped proxy geometry for the result seen in Fig-ure 1(b) is given in Figure 5 where the raycastingstarting position values are depicted.

To change the visual representation of the ray-casted structures, we use style transfer func-tions [2]. Style transfer functions utilize light-ing information as acquired from orthogonally pro-jected lit spheres. This technique makes it simple toachieve view-aligned lighting. With a style trans-fer function it is easy to switch from simple dif-fuse Lambert shading to Phong like shading by us-ing different lit spheres. This technique enables usto simulate the non-photo realistic illustration styleof medical anatomy illustrations like the ones bySobotta [16]. See Figure 7 for examples.

We also need a mechanism to control what stylesthe raycaster should use for a specific substructure.Our solution to this is to use raycastingportals.Usually for GPU accelerated volume raycasting, a

Figure 5: A visualization of rendered proxy geome-tries with the starting positions for the raycasting.

full-screen quad is rendered to the screen which isalso a trigger for the GPU program to generate pix-els. Instead of creating a full-screen quad we rendera quad for every structure that we need raycastingfor. A quad is centered on every node and the sizeof the quad is equal to the bounding square of thenode drawn at that node. We call these quads ray-casting portals. The main feature of these portals isthat we can now communicate portal-specific ren-dering parameters to the GPU program via a shaderuniform variable. The uniform variable contains alist of styles for the structures that should be visibleand highlighted, and a zero reference to all struc-tures that should be invisible.

To further increase the amount of informationthat is associated with the hierarchical elements andtheir volumetric nature, we have additionally im-plemented axis-aligned slicing. This visualizationgives the user the opportunity to study the under-lying MRI data and not just the raycasting of thestructures. Figure 6 illustrates the combination of3D and 2D information. Some of the possible com-binations of slicing and volume rendering are illus-trated in Figure 6.

3.4 Hierarchical Styling

The hierarchical organization of the data is usedwhen we convey the hierarchical arrangement be-tween sub-volumes. We change the visual repre-sentation of structures to clearly illustrate the hier-archical relationships. We have implemented twoways of interacting with the visual representationof structures to achieve two different goals. The

(a) (b) (c) (d)

Figure 6: Axis-aligned slicing of the frontal lobe: (a) direct view of a slicein the z direction. (b) structureand slice, features in front of the slice have been removed. (c, d) structure and slice from two differentview-points (no removal of structures).

first goal is to visualize the hierarchical relation-ship between the parent and the child. This is doneby showing how the parent node is composed ofthe child nodes by uniquely coloring each structure.The second goal is to show how a single structure ormultiple structures are spatially located in all appli-cable hierarchy levels. This is achieved by the userinteractively setting the style for a structure. Thefirst approach is shown in Figure 1(a) and the sec-ond one is shown in Figure 1(b).

Our goal is to provide the user with multipleways of seeing how the hierarchical structures areorganized. One way of doing this is to color eachstructure uniquely so that they are easy to differ-entiate from each other. When the mouse hoversover a node, a feature called auto-styling is initiated.Auto-styling is the visual result of applying pre-defined style transfer functions to structures. Thisfeature simplifies the navigation through the hierar-chy. In addition it enhances the mental image theuser has of the 3D structures of the brain. We havedefined eight perceptually different styles based onpastel versions of red, green, blue, cyan, magenta,yellow, orange, and a darker blue. These colorshave been selected based on contrast and lightness.The pre-defined styles are customizable.

If the mouse hovers over a group node, all itschildren are set to one of the predefined styles. Thechildren are assigned styles in the same order asthey are defined in the hierarchy. The style appliedto a child is also used in the visual representationof the group node. This can be seen in Figure 1(a)and in Figure 8(a). In Figure 7 it is possible to seewhere the right medial temporal gyrus is located inthe right temporal lobe using this technique.

If the mouse hovers over a child node, this node

is displayed using the same pre-defined style as inthe previous case, but the group node is displayeddifferently. The group node is displayed using thegroup’s default style (defaults to grey). The selectedstructure is the only structure that is displayed witha different visual representation. This can be ob-served in Figure 8(b) where the selected structurecan be seen in the group in light blue color.

Tracking a specific sub-structure in several dif-ferent hierarchical levels is also possible. The usercan change the style of a structure and then moveup in the hierarchy to observe where the structureis located relative to higher levels in the hierarchy.Figure 9 illustrates this visual enhancement. Theuser has changed the style of the medial gyrus partof the temporal lobe. Moving up the hierarchy, thestructure is now highlighted with the user selectedorange style.

3.5 Interaction and Navigation

We enable the user to navigate by panning the data,hierarchically guided navigation, seamless zoomingfrom the overview to the volume rendering, and byrotation of volume geometry. The user can hoverover nodes to initiate auto-styling or manually setthe style of interesting structures so they can eas-ily be tracked. When exploring the data, the usercan move the entire view or zoom in and out. Ex-ploring the hierarchy in a zoomed-in manner can bea tedious task. A lot of dragging with the mouseis necessary to move around. Therefore we haveimplemented guided navigation that helps the userto move between nodes. If the user selects a node,the view centers on that node. This also meansthat zooming in and out is now centered on this

Figure 7: Auto-styling of hierarchically linked structures. From left to right: entire brain, right hemisphere,right cortex, right temporal lobe and right medial temporal gyrus.

(a) (b)

Figure 8: (a) Auto-styling of nodes in a leaf-cluster when the group nodeis selected. (b) Auto-styling ofnodes in a leaf-cluster when the leaf node is selected.

node. If the node is a leaf node then the guidednavigation will move the focus to the parent node.If the user initiates guided navigation on a repli-cated group node then the focus changes betweenthe group node and its replica.

When the user is at an abstraction level, whichshows volume rendered structures, rotation of thestructures is possible. Rotation applies to all struc-tures so that there is always a coherency betweenthe views of the volumetric data. The user interfacealso lets the user change the visual representation ofa selected structure so that it is possible to see wherea feature is located in higher level structures.

4 Results

Seamless zooming from the contextual overviewdown to the data at the lowest hierarchical levelcan be observed in Figure 1 and Figure 4. In Fig-

ure 4(a) the complete overview is shown. The leftpart of this overview represents the left hemisphereand the right part represents the right hemisphere.Figure 4(b) is the left hemisphere only. At thislevel it is possible to see thumbnail sized volumeraycastings of the structures. As the labels on theright hemisphere in the figure indicate, the upperpart is the cortex and the lower part represents thesub-cortical areas. Zooming into the dashed orangerectangle we get the information as shown in Fig-ure 1. In this figure we start to see details of thevolume renderings and labels are readable. It is forexample possible to see where the lingual is locatedin the occipital lobe (green structure in Figure 1(a)).

During inspection of the volume data at the low-est level of data exploration, it is possible to en-hance the volume rendering with slicing so that theoriginal data might be inspected. Figure 6 showsthis concept. The user can choose to only show the

Figure 9: User-specified styling of a selected structure. From left to right: right medial temporal gyrus,right temporal lobe, right cortex, right hemisphere and entire brain.

slices. It is also possible to mix the two types ofvisualization and in this way get a higher level ofunderstanding of the 3D nature of structures rela-tive to 2D slices.

Auto-styling is highlighted in Figure 7. This Fig-ure shows how the styling looks like at the differenthierarchy levels if a user would start at the brain andnavigate down to the medial gyrus of the temporallobe. Going the other way, from the gyrus, and see-ing how this structure is positioned relative to thehigher level structures is illustrated in Figure 9.

We have implemented our approach in Java us-ing OpenGL on a GeForce 8800 GTS. We used theOpenGL Shading Language extensions, texture ar-rays and integer texel look-ups, provided in the Sha-derModel 4.0 specification.

The Bert data-set that we have visualized tocreate the results in this work is T1-weighted 3DMRI data at an isotropic resolution of 2563 with1 µL (microliter) voxels. This data-set has beenrun through the automated segmentation workflowof FreeSurfer. The surface reconstruction and vol-ume labeling has produced approximately 80 sub-cortical and 68 cortical segmentations. We have inaddition created a data-set that describes the hier-archical semantics of the segmentations. Based onthese three data-sets we have generated a contextualoverview that illustrates the hierarchy of the data.We render the MRI data using volume raycasting.

The context view without any volume raycastingrenders at around 50 frames per second. Addingvolume raycasting reduces the rendering speed toaround 10 frames per second. The performance de-pends on how many pixels a structure covers, thenumber of voxels in the rendered structure and theamount of transparency defined in the style transferfunction. To keep the interaction fast the raycast-ing is speeded up during interaction by doubling thesampling distance. This effectively doubles the ren-dering speed.

5 Discussion and Future Work

The presented approach to hierarchical brain visu-alization has an immediate use in education. Learn-ing neuroanatomy and neurophysiology has alwaysbeen regarded as challenging for medical students.This is due to the inherently complex and abstractstructure of the nervous system, and to the intricatethree-dimensional organization of the human brain.Our approach allows to visualize how the brain’scomponents fit together, both in a strictly anatom-ical setting, and also in a functional-hierarchicalmanner. A daunting task is to understand whichparts of the brain are in connection with which oth-ers and how they function together. The here pre-sented approach has the potential to increase the ef-ficiency of learning and ease the process of com-prehension. Furthermore in radiology, students arerequired to relate planar images in different orien-tations to recorded volumes from modalities suchas Computed Tomography and MRI. Our approachprovides a way to investigate how the cut-planes ac-tually represent parts of the volume. This will helpstudents to understand the process of planar imag-ing, which is something many find difficult, andhelp themselves relate the two-dimensional imageto the actual volume in real-time.

In a clinical setting it is important to conveysubject-specific visualizations from the data. In de-mentia research, for example, it is of interest tocompare volumetrics between a statistically normalbrain and probable dementia patients. This compar-ison can be implemented in the hierarchical visual-ization to draw attention to structures that deviate,or not, from the normal population. We also fore-see that including data obtained from other imagingsources such as functional MRI and diffusion ten-sor imaging can have a potential use in a clinicalsetting. Further research and evalutation of theseapproaches is left to future work.

6 Conclusion

We have presented an approach to visualize hier-archical volume data and enable hierarchy-basedinteraction. We have described how to generatethe context view that enables a seamless navigationfrom the abstraction to the data. The visualizationof structures in focus is automatically changed toreflect the hierarchical nature of the data.

Based on feedback from the medical side, ournew visualization concept is considered promisingand useful for medical education, especially forteaching radiologists. They usually look at 2Dslices only and it is very hard for students to graspthe 3D structure of the structures observed on planarslices. With our approach they can select a portionof the data according to the hierarchical structure,and avoid that they are overloaded with the entiredata-set. The integrated slice rendering gives thestudent a correspondence between 2D and 3D.

In general we do not see any difficulties to adaptour approach to other hierarchical structures suchas the bones of the human hand. We think that thereare other scientific domains with data containing hi-erarchies that would benefit from such an approach,also. An outlook for the future in hierarchical brainvisualization comes from our medical collaborators.They observe that the anatomical structure of thebrain is rather artifical and in many cases does notfit the functional hierarchy. Visualizing functionalhierarchies will move the applicability of this toolfrom educational use to clinical use.

References

[1] K. Andrews and H. Heidegger. Informationslices: Visualising and exploring large hierar-chies using cascading, semi-circular discs. InLBHT, IEEE InfoVis ’98, pages 9–12.

[2] S. Bruckner and M. E. Groller. Style trans-fer functions for illustrative volume rendering.CGF, 26(3):715–724, Sep 2007.

[3] K. Engel, M. Hadwiger, J. M. Kniss, C. Rezk-Salama, and D. Weiskopf.Real-time VolumeGraphics. A. K. Peters, 2006.

[4] Freesurfer. http://surfer.nmr.mgh.harvard.edu.[5] I. Herman, G. Melancon, and M. S. Marshall.

Graph visualization and navigation in infor-mation visualization: A survey.IEEE TVCG,6(1):24–43, 2000.

[6] D. Holten. Hierarchical edge bundles: Visu-alization of adjacency relations in hierarchicaldata.IEEE TVCG, 12(5):741–748, Oct 2006.

[7] T. Keahey. The generalized detail-in-contextproblem.IEEE InfoVis ’98, pages 44–51.

[8] J. Kruger and R. Westermann. Accelerationtechniques for GPU-based volume rendering.In IEEE Visualization 2003, pages 287–292.

[9] G. Melancon and I. Herman. Circular draw-ings of rooted trees. Technical Report INS-R9817, CWI, Amsterdam, Netherlands, 1998.

[10] T. Munzner. H3: laying out large directedgraphs in 3D hyperbolic space.IEEE InfoVis’97, pages 2–10.

[11] P. Neumann, S. Schlechtweg, and M. Carpen-dale. Arctrees: Visualizing relations in hierar-chical data. InEuroVis ’05, pages 53–60.

[12] A. Pommert, Schubert, Riemer, Schiemann,Tiede, and Hohne. Symbolic modeling of hu-man anatomy for visualization and simulation.In Vis. in Biomed. Comp., volume 2359, pages412–423. SPIE, 1994.

[13] G. Robertson, J. Mackinlay, and S. Card.Cone trees: animated 3D visualizations of hi-erarchical information. InProc. CHI ’91,pages 189–194. ACM.

[14] B. Shneiderman. Tree visualization with tree-maps: 2-d space-filling approach.ACM Trans.Graph., 11(1):92–99, 1992.

[15] B. Shneiderman and M. Wattenberg. Orderedtreemap layouts.InfoVis ’01, pages 73–78.

[16] J. Sobotta.Sobotta Atlas of Human Anatomy.Lippincott Williams & Wilkin, 2001.

[17] J. Stasko and E. Zhang. Focus+context dis-play and navigation techniques for enhancingradial, space-filling hierarchy visualizations.IEEE InfoVis 2000, pages 57–65.

[18] S. T. Teoh and K.-L. Ma. RINGS: A techniquefor visualizing large hierarchies. InGD ’02,pages 268–275.

[19] Y. Wang, S. Teoh, and K.-L. Ma. Evaluat-ing the effectiveness of tree visualization sys-tems for knowledge discovery. InEuroVis ’06,pages 67–74.

[20] J. Wijk and Nuij. Smooth and efficient zoom-ing and panning.InfoVis ’03, pages 15–23.

[21] J. Yang, M. Ward, and E. Rundensteiner. Inter-ring: an interactive tool for visually navigat-ing and manipulating hierarchical structures.IEEE InfoVis 2002, pages 77–84.