`
`A Hybrid User Interface for Manipulation of Volumetric Medical Data
`
`Conference Paper · April 2006
`
`DOI: 10.1109/VR.2006.8 · Source: IEEE Xplore
`
`CITATIONS
`46
`
`5 authors, including:
`
`READS
`275
`
`Alexander Bornik
`Ludwig Boltzmann Institute for Archaeological Prospection and Virtual Archaeology
`
`Bernhard Reitinger
`Robotic Eyes
`
`42 PUBLICATIONS 739 CITATIONS
`
`SEE PROFILE
`
`55 PUBLICATIONS 1,380 CITATIONS
`
`SEE PROFILE
`
`Dieter Schmalstieg
`Graz University of Technology
`
`539 PUBLICATIONS 13,252 CITATIONS
`
`SEE PROFILE
`
`Some of the authors of this publication are also working on these related projects:
`
`HYDROSYS - Mobile environmental monitoring tools View project
`
`Training and Workflows in Mixed Reality View project
`
`All content following this page was uploaded by Alexander Bornik on 05 June 2014.
`
`The user has requested enhancement of the downloaded file.
`
`0001
`
`Align EX1023
`Align v. 3Shape
`IPR2022-00145
`
`
`
`A Hybrid User Interface for Manipulation of Volumetric Medical Data
`
`Alexander Bornik∗
`
`Reinhard Beichel
`
`Ernst Kruijff
`
`Bernhard Reitinger
`
`Dieter Schmalstieg
`
`Institute for Computer Graphics and Vision
`Graz University of Technology
`
`ABSTRACT
`
`This paper presents a novel system for interactive visualization and
`manipulation of medical datasets for surgery planning based on a
`hybrid VR / Tablet PC user interface. The goal of the system is to
`facilitate efficient visual inspection and correction of surface mod-
`els generated by automated segmentation algorithms based on x-ray
`computed tomography scans, needed for planning surgical resec-
`tions of liver tumors. Factors like the quality of the visualization,
`nature of the dataset and interaction efficiency strongly influence
`system design decisions, in particular the design of the user inter-
`face, input devices and interaction techniques, leading to a hybrid
`setup. Finally, a user study is presented, which characterizes the
`system in terms of method efficiency and usability.
`CR Categories:
`C.2.4 [Computer-Communication Networks]:
`Distributed Systems—Distributed Applications; I.3.5 [Computer
`Graphics]: Computational Geometry and Object Modeling—
`Boundary Representations; I.3.7 [Computer Graphics]: Three-
`Dimensional Graphics and Realism—Virtual Reality; J.3.2 [Med-
`ical information systems]: Project and People Management—Life
`Cycle
`
`1
`
`INTRODUCTION
`
`Imaging modalities like X-ray computed tomography (CT) or
`magnetic resonance tomography (MR) are important information
`sources for surgical planning. Proper planning requires physicians
`to understand the 3D relations within the dataset. For example, the
`resection of liver tumor requires understanding the arrangement of
`liver tissue, vasculature and tumor. Looking at individual 2D slices
`of CT data using conventional radiological workstation software
`makes this task difficult. Our aim was therefore to build a system
`for liver surgery planning using Virtual Reality (VR) techniques,
`capable of supporting radiologists and surgeons by providing vi-
`sualization of 3D medical models and tools for computer-assisted
`planning of the surgical intervention.
`Typically, the first step in liver surgery planning is segmentation
`of the individual structures, needed to plan the surgical interven-
`tion. This task can be done manually, but this is tedious and time
`consuming, since it involves drawing contours on several hundred
`slices.
`However, a fully automated segmentation of the liver is difficult
`to achieve, because the shape of the human liver highly varies. This
`fact makes it almost impossible to use a priori shape knowledge for
`the design of a segmentation algorithm. In addition, the gray-value
`appearance can show large variations due to pathological changes
`of the liver, which can cause problems in distinguishing the liver
`from adjacent organs with similar gray-values (e.g. heart or colon).
`Furthermore, tumors located close to the liver boundary might be
`excluded from the segmentation.
`
`∗e-mail:bornik@icg.tu-graz.ac.at
`
`In advanced automatic segmentation algorithms, the segmenta-
`tion problems are usually limited to local errors, while most areas of
`the liver boundary can be correctly found using the automatic algo-
`rithms. A radiologist’s task can therefore be simplified from manual
`contour specification to interactively correcting errors in segmented
`datasets. This segmentation refinement approach is expected to be
`much less time consuming in most cases.
`At a first glance 3D segmentation refinement tools afford VR
`techniques: Stereoscopic visualization provides good 3D percep-
`tion of the dataset, whereas tracked input devices allow for direct
`3D interaction with the dataset. However, 2D screens have a much
`higher resolution than their 3D counterparts, and an inexpensive op-
`tical mouse easily outperforms high-end tracking devices in terms
`of accuracy when precision input in 2D is required. In the medical
`field, where imprecision may have dire consequences, the virtues of
`established 2D techniques should not be discarded lightly. More-
`over, physicians are used to desktop interfaces, and in particular for
`system control, VR interfaces are not yet mature.
`These considerations lead us to the design of a hybrid user in-
`terface that combines multiple display and interaction techniques,
`in order to match the work processes at hand. The objective of the
`hybrid user interface is to pair 3D perception and direct 3D interac-
`tion with 2D system control and precise 2D interaction. For such an
`interface, it is important that the flow of action of working between
`2D and 3D visualization and interaction techniques is not disturbed.
`Both the different views and the interaction with the data need to be
`handled coherently.
`To ease the transition between the interface modalities, a hybrid
`input device, which can be conveniently used in all 2D and 3D
`tasks, was designed and developed. A focus was put on analyz-
`ing the differences between action performance in the 2D and 3D
`domain, leading to a more extensive human factors study. This pa-
`per presents results on the complexity of tasks and their associated
`tools, and the duration of usage in reflection to ergonomics.
`
`2 RELATED WORK
`
`The three-dimensional nature of surgery planning and surgery sim-
`ulation has led researchers to the use of VR techniques. An
`overview of VR systems and human interface issues in the medi-
`cal context can be found in [21] and [9]. Liver surgery planning in
`particular has been addressed by a number of groups, although VR
`aspects are rudimentary in most projects.
`The German cancer research center (DKFZ) located in Heidel-
`berg has developed a computer-aided planning system for liver
`surgery [20]. Research was focused on medical image process-
`ing tasks such as segmentation. There have also been attempts to
`deal with segmentation errors described in [16], although the VR
`aspects of the system are limited, since all image segmentation and
`planning procedures are performed on a normal desktop PC.
`The Center for Medical Diagnostic Systems and Visualization
`(MeVis) in Bremen has developed a desktop-based liver surgery
`planning system. The main research focus was segmentation and
`modeling of liver structures [15].
`Researchers at INRIA have addressing several aspects of liver
`surgery planning, such as the segmentation of the liver surface using
`
`0002
`
`
`
`deformable surface models. Later the group worked on surgical
`simulation with realistic liver tissue models using force feedback
`input devices [8]. The virtual liver surgery planning system [3],
`which is the foundation of the work presented in this paper, has
`been developed at Graz University of Technology since 2000. An
`earlier version of the VR based segmentation refinement toolset was
`described in [4].
`Segmentation refinement is a rather new field and there are
`hardly any publications dealing directly with it. 2D segmentation
`can be trivially implemented as a painting tool, but this approach
`is ineffective for large 3D datasets. Interactive segmentation tech-
`niques can be seen as closely related, although they address the
`segmentation problem and not the correction of erroneous segmen-
`tation. An interactive segmentation approach named Live-Wire was
`introduced in [1]. It reduces the amount of user interaction required
`to segment the object boundary. An extension to 3D data can be
`found in [10]. To our knowledge, the desktop based interactive ap-
`proach described in [19] is the only method based on the segmen-
`tation refinement principle. An application in the context of data
`preparation for liver surgery was reported in [2].
`3D interaction with medical volumetric datasets is a reoccurring
`topic in VR. Some input devices have been developed that specif-
`ically focus at exploring medical data, including Hinckley’s prop-
`based system [18] or the Fakespace CubicMouse [13].
`Hybrid user interfaces in general are an emerging research field
`[11]. Most developments focus at combining different visual dis-
`plays, like in mixed reality setups [23]. Only a few have focused at
`truly hybrid input, such as the Virtual Tricorder [27], the Pick-and-
`Drop approach by Rekimoto [22], and some tangible user interfaces
`[26]. Handhelds and touch screens have been integrated into im-
`mersive environments for interaction purposes, like [12], [27], [14]
`and [6]. These handhelds are mostly used for GUI-style control
`elements (system control), only little work is done on direct manip-
`ulation of immersive data. Similar to handhelds, tablet interfaces
`have been designed for the display of 2D data and menus on a flat
`surface, for example [25]. Finally, a frequent approach to system
`control in immersive environments are pen-like devices, such as
`the Stylus products from Polhemus or Intersense. For an extensive
`overview of pen devices, please refer to [5].
`
`3 HYBRID USER INTERFACE
`
`Figure 1: Hybrid Setup: camera of the optical tracking system (1),
`Tablet PC and Eye of Ra (2), stereoscopic large screen projection
`system (3).
`
`s
`
`Figure 2: Desktop Setup: Tablet PC with conventional 2D User
`interface for system control; 2D view for viewing and interaction
`with the dataset using the Eye of Ra input device, which behaves
`similar to a conventional stylus in the desktop setup.
`
`3.1 Hardware Setup
`
`The hardware setup consists of two main parts, the VR system and
`the 2D system. The VR system’s display is a large stereo wall
`(stereoscopic back projection screen, 375cm diameter, 1280x1024
`pixels) viewed with shutter glasses. A Barco Galaxy 3-chip DLP
`projector provides high quality active stereo rendering with very
`good channel separation, which is important when displaying vir-
`tual objects close to the user. The stereo wall is driven by a PC
`workstation (dual 3GHz Xeon, NVidia Quadro FX 3400). Optical
`tracking of the user’s head and the input device is done using an
`4-camera infrared system from Advanced Realtime Tracking.
`The desktop system is a Tablet PC (Toshiba Port´eg´e M200, 1.8
`GHz CPU, GeForce Go 5200 graphics card, 12-inch TFT touch-
`screen at 1400x1050 pixels). The Tablet PC is placed on a desk ap-
`proximately 2 meters in front of the screen, tilted at approximately
`60 degree for convenient readability. The user is seated at the desk
`so that both stereo wall and Tablet PC are within the field of view
`as shown in Figure 1.
`
`3.2 Hybrid Interaction
`
`When referring to hybrid interaction, it is important to differentiate
`between two approaches: serial and parallel integration. Using se-
`rial integration, 2D and 3D methods are used in a sequential order,
`one after each other. In parallel integration, 2D methods are quasi
`embedded and used directly to control and adapt the data in the
`immersive environment. In the virtual liver planning system, the
`interaction makes use of serial integration, in which the 2D and and
`the VR system are two separate systems that can be synchronized.
`The combination of 2D and 3D interactions can have consider-
`able advantages. 2D actions can be performed with relatively high
`precision, whereas 3D actions are executed at high speeds in spe-
`cific task situations. As such, a clear speed-accuracy trade-off can
`be noticed, depending on the task at hand. In that respect, the vir-
`tual liver planning application contains actions that are inherently
`2D (like contour editing or point-based segmentation refinement)
`or 3D (including visual inspection of mixed data, or approximation
`of surfaces).
`The main factor is the flow of action at macro and micro level,
`for which the mapping and mode changes of functionality, the syn-
`chronization of desktop and spatial environments and the focus of
`
`0003
`
`
`
`attention play a key role.
`At macro level, the performance of the actions is influenced
`by the work process preferences of the end-users: the radiologist
`prefers the desktop, whereas the surgeon can better work within
`a spatial setting. In order to access the functionality, an effective
`system control method is needed, that allows consistent interaction
`at both desktop and spatial environment. Redundant mapping of
`functionality is deliberately chosen, in order to support the work
`preferences of the end-users: all actions can be performed in the
`desktop and the spatial environment.
`At micro-level, one of the issues that affect the flow of action
`in an application is the effectiveness of performing mode changes.
`The high amount of functions cannot be accessed effectively by
`any of the currently available 3D system control techniques. There-
`fore, the only possible way for mode changes is to access most of
`the functions on a standard GUI-style menu on the desktop screen.
`However, to support frequent interaction loops in the immersive en-
`vironment, some functions are mapped to the input device. Specifi-
`cally, manipulation actions are mixed with visual inspection actions
`(e.g. navigation) at high regularity. Therefore both CT data move-
`ment and general camera movement actions are directly mapped to
`two buttons on the input device.
`Due to the different locations of the desktop display and the
`stereo wall in relation to the user, switching between desktop and
`spatial interaction (for example during mode change) necessarily
`results in a change of visual focus. Best practice demands that head
`rotation and focal plane difference are as limited as possible, with-
`out the desktop display occluding the stereo wall.
`In the current hardware setup, the Tablet PC is placed at a table,
`and put in a tilted angle towards the users. The user can conve-
`niently use the touch screen for selecting menu items or manipulat-
`ing objects. The user’s arm may be placed on the table to reduce
`fatigue. The table is placed at a specific distance from the stereo
`wall, so that stereo objects are viewed in a depth place that seems
`to be above or just behind the visuals viewed at the desktop screen.
`Consequently, both the angular movements of the head are limited,
`as well as the change of focus between depth planes. There are still
`field of view differences. Combining a smaller stereo wall with a
`larger touch screen in an L-shape like configuration may improve
`this issue.
`Performing the different actions in desktop or spatial mode nec-
`essarily leads to different kinds of input.
`In the spatial environ-
`ment, most actions are coarse, mixed with some more fine-grained
`actions, whereas at the desktop, all actions are fine-grained. The
`different kinds of performances, and the necessity to make use of
`a pen-like device to control the touch screen lead to different kinds
`of dynamical coupling between hand and device. This is mostly
`caused by the different kinds of grips to the device that match the
`precision needed to perform the task, as will be illuminated in the
`next section.
`
`3.3 Hybrid Input Device
`
`For the design of the new device, a close analysis of the tasks and
`the associated hand-device couplings and movements was made. In
`initial tests, a specific device for two-handed interaction was not
`found necessary because all tasks could be performed well with
`one hand. However, we observed that the device would need to
`allow for both power and precision grasps. Rotational movements
`for putting clipping plane or CT data plane are generally performed
`in high-speed and lower accuracy (sweeping task), whereas other
`tools like deformation demand lower speed and higher accuracy,
`and are better performed in precision grip.
`To get an idea of a basic form of the device, we tried to match the
`movement and rotation patterns to existing devices. It was found
`that the hand activity matches partially a flying mouse, partially a
`
`pen-like device. The pen characteristics were considered impor-
`tant, since the device needed to function as pen-input device for the
`Tablet PC.
`This resulted in an attempt to merge flying mouse and pen shapes
`into one single design, which allows for an unobtrusive switching
`between power and precision grasps. From clay models, we ar-
`rived at the shape shown in Figure 3. Due to the visual shape of
`the device, it was nicknamed Eye of Ra. The form needed to be
`large enough to enclose the electronics, which were taken from an
`EZ5 Optical Pen Mouse, which has a very small circuit board. The
`wireless connection was tested and found suitable when combined
`with a longer antenna in the device. The final device was made
`from carbon and fiberglass mats layered with epoxy glue, which re-
`sults in a lightweight yet sturdy surface. The original button casing
`from the EZ5 was directly included in the design, in order to make
`a stable connection between device casing and electronics. Finally,
`the tip of the electromagnetic pen for the Tablet PC was embedded
`at the front of the device, and retro-reflective markers required for
`tracking were rigidly mounted on the body of the device.
`
`Figure 3: Eye of Ra - Input device for the hybrid user interface: The
`tip contains a conventional Tablet PC stylus tip for 2D interaction.
`Two buttons and a scroll wheel are used to trigger 3D interaction
`tasks. It is equipped with retro-reflective targets for optical tracking.
`The device is connected to the Tablet PC via a cable. Note the two
`different ways of grasping the device, the power grasp on the left and
`the precision grasp on the right.
`
`The shape of the hybrid interaction device allows for easy
`switching between flying mouse and pen mode. By pronating the
`forearm, and slightly changing the position of the fingers (mostly
`moving the thumb), the user can easily change between the differ-
`ent modes. This allows for dynamic coupling between device and
`hand without the user actively noticing it.
`
`3.4 Software
`
`Following the overall hybrid approach, the software of the system
`consists of two collaborating applications, a desktop application
`and a VR application. Both parts of the system are closely cou-
`pled and share a large portion of the code. The dataset is visualized
`on both systems simultaneously. Interaction with the data can take
`place in either application, while system control tasks like loading
`the datasets or setting parameters are limited to the 2D menu system
`on the Tablet PC. All user interaction is performed using the Eye of
`Ra input device described in Section 3.3, which acts like a normal
`Tablet PC stylus in the desktop application, while 6-DOF tracking,
`a scroll wheel and buttons on the device are used to trigger input in
`the VR application.
`Both applications involve 3D rendering based on Coin1, a scene
`graph library compatible to the Open Inventor standard. The desk-
`
`1http://www.coin3d.org
`
`0004
`
`
`
`top application uses the Qt 2 framework for graphical user interface
`programming. The VR application is based on the Studierstube [24]
`VR/AR library, which builds on top of Coin and provides handling
`of VR devices such as stereo wall and tracking as well as convenient
`programming of 3D interaction with the scene graph.
`
`Hybrid System
`
`Desktop System
`
`Desktop Application
`
`Application
`Layer
`
`VR Sytem
`
`VR Application
`
`Studierstube
`
`Studierstube
`
`Coin
`
`Coin
`
`Toolkit
`Layer
`
`DIV
`
`DIV
`
`Graphics
`Hardware
`
`Network
`Hardware
`
`Hardware
`Layer
`
`Network
`Hardware
`
`Graphics
`Hardware
`
`Figure 4: Software Architecture: The two separate applications share
`data via a network connection based on a distributed scene graph.
`
`Synchronization between the desktop and VR application is
`based on a distributed shared scene graph extension called
`DIV [17]. The two applications share synchronized copies of the
`scene graph, which stores all geometric and application relevant
`data. Modifications to one copy of the scene graph will be propa-
`gated to the other copy, and vice versa. The synchronization hap-
`pens automatically within the scene graph library, and need not be
`managed by the application programmer explicitly. Figure 4 gives
`an overview of the system architecture of the hybrid system.
`
`4.1 Model Inspection
`
`The first step, model inspection, can be performed on the Tablet PC
`screen using the Eye of Ra’s tip for interaction with the rotation,
`movement and scaling controls in the 2D user interface. In VR the
`model can be moved and rotated by pressing the scroll-wheel button
`on the input device, which fixes the model to the input device, while
`moving the device. The model navigation feature is bound to the
`scroll-wheel and is permanently available.
`CT data is visualized on a 2D cutting plane that can be arbitrarily
`placed inside the CT scan volume. On the Tablet PC the plane can
`be manipulated by dragging 3D control widgets provided by the
`scene graph library. In the VR system the cutting plane, visualized
`as a rectangle attached to the input device, can be set by dragging in
`3D with a specific button pressed. Like the model transformation,
`the cutting plane manipulation feature is permanently available.
`The user may also configure most visualization parameters. For
`example, they can choose to show the surface model in any com-
`bination of wireframe, Gouraud shading and textured with the CT
`data. An optional plane clipping the 3D model slightly above the
`cutting plane allows to inspect the surface model near the clipping
`plane more efficiently.
`
`4.2 Error Marking
`
`For efficient organization of the correction procedure, the user may
`mark regions according to the type of observed error by painting
`a ”traffic light” color code - green, yellow or red. Green indicates
`that a portion of the surface is correct and will be immutable by sub-
`sequent correction operations. Yellow indicates that the surface is
`mostly correct but may be moderately altered from its current state
`as needed for example to smooth out differences at region bound-
`aries. Finally, red indicates the surface is incorrect and may be
`drastically altered by the error correction tools. The marking is
`done by painting on the surface either on the Tablet PC or in the
`VR environment with a brush of adjustable size.
`
`4
`
`INTERACTION TOOLS
`
`4.3 Error Correction
`
`There are three main tasks necessary to improve a segmented sur-
`face model:
`
`• Model Inspection - The user tries to locate errors in the sur-
`face model by comparing raw CT data to the boundary of the
`surface.
`
`• Error Marking - Regions of the surface model that were
`found erroneous in the inspection step are marked for further
`processing. This allows to restrict the following correction
`step to the erroneous regions, and avoids accidentally modi-
`fying correct regions.
`
`• Error Correction - Marked regions are corrected using spe-
`cial correction tools based on mesh deformation.
`
`Usually these tasks are not performed in a strict order. For ex-
`ample, model inspection is repeatedly required throughout the cor-
`rection task. Individual corrected parts should be marked as final
`after successful correction.
`The following sections will give an overview of the functionality
`of the system and the tools provided, mostly from the user’s point
`of view, while technical details of the implementation of the seg-
`mentation refinement tools will be described in Section 5.2.
`
`2http://www.trolltech.no/products/qt
`
`The presented system allows for correction of segmentation errors
`using a number of different tools for interactively deforming the
`surface representation of the object.
`The sphere deformation tool consists of a sphere of user-defined
`radius which can be interactively placed in the datasets. In the VR
`system this is moved in place by moving the input device, while
`the tool position in the desktop setup is calculated as the position
`on the cutting plane corresponding to the 2D position of the cursor.
`Triggering sphere deformation causes object surface parts located
`within the sphere shape to be successively moved out of the sphere
`on the shortest possible path. Therefore, placing the sphere tool
`so that most parts of it are outside the object, causes its surface to
`move inwards, while outward movement is achieved by placing the
`sphere mostly inside the object. Moving the input device, while the
`deformation tool is active, causes the tool to respond, just like if one
`was deforming a piece of clay using a real world modeling tool.
`The plane deformation tool is much like the sphere deformation
`tools, except that its behavior is similar to modeling using a scraper.
`It can be used to flatten the object’s surface. In the VR system the
`position and orientation of the tools is directly determined by the
`input device, while cutting plane and the pen stroke direction define
`the tool’s behavior in the desktop setup.
`Fine grained deformation can be achieved using the point drag-
`ging tool, which can be used to pick individual surface vertices and
`move them directly to the desired location, while the surface de-
`forms like a rubber sheet in the vicinity. Figure 5 shows a screen-
`shot of the tools described above.
`
`0005
`
`
`
`graphics memory in each frame.
`Painting the mesh surface with the error marking tools intro-
`duced in Section 4.2 does not only alter the surface color. Paint-
`ing the mesh in red affects the deformable model in a way, that
`forces towards the erroneous segmentation boundary are not calcu-
`lated anymore. The mesh does not deform in these regions until
`until refinement tool induced forces apply. If the mesh is painted
`green, indicating that the surface is correct, no forces are calculated
`for the affected vertices. The standard color is yellow. In this case
`the forces towards the segmentation apply. Regularizing forces may
`still cause limited response to nearby tool based deformation.
`Applying the sphere deformation tool presented Section 4.3 re-
`sults in force calculation for all red colored surface vertices located
`within the tool’s influence range. The force vector direction is
`spanned by the sphere center and the affected vertex. The length
`of the vector is the distance from the affected vertex to the sphere
`border. The plane deformation tool works similarly.
`The point dragging affects the vertex closest to the input device
`location when it is triggered. The vertex can be placed freely in
`space. It is immediately removed from the set of active vertices,
`but its three neighbors are added to the active set, in case they are
`either red or yellow. When the vertex is released it keeps its current
`position, while the red of yellow adjacent parts of the mesh deforms
`and/or restructures until all quality criterions have been met.
`
`6 EVALUATION
`
`6.1 Methodology and Procedure
`
`The overall goal of the evaluation is to investigate the validity of
`our hybrid interface for liver surgery planning by comparing spa-
`tial (3D) and constrained (2D) tasks. Therefore, the evaluation is
`planned in two steps:
`
`1. The first step examined the general spatial manipulation tools
`described in Section 4.
`
`2. In the second step specific constrained tasks like segmentation
`refinement based on local contour drawing will be evaluated.
`
`In this paper, we only report on the first evaluation step. The
`different modes of the system, namely desktop, spatial interaction,
`or hybrid mode were tested in a comparative study. The evalua-
`tion included mostly empirical testing, with some analytical meth-
`ods, using a variety of data collection methods. The evaluation ad-
`dressed several issues relevant in complex interaction tasks, in par-
`ticular learning curve effects and mode switching. The evaluation
`included both qualitative and quantitative components, in which the
`user attitude and psycho-physiological abilities were collected and
`analyzed. All results were cross-compared to see if there were any
`notable differences between the user attitude towards the system,
`and the data collected through observation and recording.
`The qualitative component of the evaluation was dominated by
`the subjective measurements obtained from the questionnaires, and
`the quasi thinking aloud protocols. The thinking aloud protocol
`was more or less a notification of the thoughts that were expressed
`by the subjects. The subjects were asked to speak, but not forced.
`As such, results from the thinking aloud protocols differed between
`users, since expression levels differed between persons. The ques-
`tionnaire was mostly focused at the user satisfaction, by validating
`17 questions. The main factors included were user learning curve,
`attitude towards tools, easiness and effectives of tools, user com-
`fort including fatigue and device ergonomics, and attention. Hence,
`the questionnaires focused at the main issues specified in the hybrid
`interaction methodology applied, as described in Section 4.
`The quantitative data were collected from the external observer’s
`notes, the quality of the final liver model delivered by the subjects,
`
`Interaction Screenshot: Deformation of the model using
`Figure 5:
`the sphere deformation tool on the Tablet PC. The correctness of
`the model can be verified using the cutting plane chowing CT data.
`Correct and erroneous regions have been marked before.
`
`5 METHODS
`
`5.1 Cutting Planes
`
`The 2D cutting planes are based on OpenGL 3D textures derived
`from the CT data by downsampling to fit the texture volume into
`the graphics card memory (2563 in the VR setup and 1283 on the
`Tablet PC). The 3D texture is displayed while the plane is interac-
`tively moved. Once the user has fixed the plane’s position, a 2D
`texture for the plane is sampled at the full resolution of the dataset,
`delivering maximum quality. Plane sampling is decoupled from
`rendering, since it would impact the frame rate.
`
`5.2 Surface Model
`
`The presented tools are all based on deformable simplex
`meshes [7]. A simplex mesh is a special surface mesh with the
`property, that each vertex has exactly three neighbors, which makes
`it easy to calculate surfaces properties such as curvature and conse-
`quently to set up a deformable model based on a Newtonian law of
`motion, involving regulating forces, and other forces to deform the
`mesh. In order to avoid mesh degeneration, polygon splitting and
`merging operations are performed based on mesh quality criteria.
`In our system we formulate forces towards the boundary of the
`binary segmentation calculated using an automated segmentation
`approach. This leads to a mesh accurately representing the seg-
`mentation.
`In a VR setup the frame rates must be high, in oder not to dis-
`comfort the user, while the model should still be accurate. On aver-
`age this leads to simplex meshes of around 250.000 polygons. We
`employ vertex buffer objects available on newer graphics hardware.
`They allow for mapping graphics memory into main memory which
`makes selective updates on the graphics card possible.
`For efficiency, we keep a set of active vertices. When a vertex
`does not move significantly over several iterations, it is removed
`from this set and only added to it again, when neighboring vertices
`move significantly or new forces are applied. The refinement tools
`described in Section 4 alter the mesh locally, so the set of active
`vertices is always much smaller than the total number of vertices.
`Note, that only vertices altered in an iteration need to be updated in
`
`0006
`
`
`
`and the logging files that tracked duration and changes of inter-
`action modes. The observer noted all question asked by the user
`as well as user behavior (grasps, observable dexterity in fingers
`and wrist, arm-hand steadiness, attention to desktop and projection
`screen). Furthermore, the work-flow was observed and later com-
`pared with logs. Finally, a comparison was performed between the
`data produced by the subjects, and the best-practice model provided
`by an expert user.
`The different steps of the evaluation were as follows: Subjects
`were first introduced to the system by the instructor, taking a 10
`minute tour through the software. After the introduction, users
`could make use of the system for 12 minutes and ask questions.
`Next, the tests were performed. Use