Incorporating Visible Human Project Data into the Undergraduate Anatomy&Physiology Curriculum

Steven Senger
Department of Computer Science
Univ. of Wisconsin - La Crosse
senger@csfac.uwlax.edu
http://www.visu.uwlax.edu

Abstract

The University of Wisconsin - La Crosse is developing a software environment for working with the Visible Human Data Set to be used by undergraduates in anatomy and physiology. The software environment provides students with a "personal digital cadaver" for study. The system incorporates a volume rendering daemon for imaging the digital cadaver. The system also includes the concept of an anatomical notebook in which students record and annotate their studies.

1. Introduction

The University of Wisconsin - La Crosse (UW-L) has begun a project to enhance instruction in the pre-professional allied health science curricula by employing the anatomical data made available by the National Library of Medicine's Visible Human Project (TM) [4]. The Visible Human Data Set (VHD) has produced approximately 16GB of raw data consisting of full body MRI and CT scans and high resolution color cryosection images of a male cadaver. The availability of this data has the potential to offer anatomy students an information resource of unprecedented detail and flexibility. By virtue of its size and raw form the VHD requires appropriate processing before information can be effectively extracted. It is, however, exactly these attributes which give it such a broad range of potential applications. Anatomical drawings and models are all constrained in their ability to visualize. They carry the imprint of the ideas which went into their creation, making it is impossible to extract information which was not explicitly placed in them. The VHD carries no such inherent constraints or preconceived uses. The VHD can potentially support visualization of any structure present, limited only by the resolution of the data set.

The UW-L effort is directed at using the VHD as more than just a repository of prepared images to be used in lectures or as textbook supplements. Rather the goal of the project is to provide students with an interactive, investigative environment in which they control the tools for processing and imaging the data in the course of their study of human anatomy.

2. Personal Digital Cadavers

Students at the undergraduate level typically have little or no individual access to physical human cadavers. At UW-L students are limited to viewing, in a laboratory session, a single cadaver prepared in advance by the instructor. This places students in a primarily passive role, observing the anatomical detail which someone else has worked to expose. There is little question that this passive role is disadvantageous in every regard and forms an obstacle to internalizing the anatomical information presented. Learning is best reinforced when the learner is an active participant and is required to build upon existing knowledge in order to discover new information.

Given that individual access to cadavers is prohibitive, the UW-L project proposes to give each student a "personal digital cadaver". This term is intended to suggest a software environment which supports

Provided with such an environment students would assume primary responsibility for their examination of anatomical data. With this responsibility would come the requirement that they determine for themselves when they have produced a useful presentation of a structure. By taking an active role in identifying and revealing anatomic structures students would necessarily utilize existing knowledge in determining how to best exhibit new features and their surrounding structures. This cycle of reinforcing and expanding knowledge through reuse and adaptation is the foundation of all good pedagogy. In this environment the images and descriptions they produce will carry their imprint, and will exhibit the information and understanding they have gleaned from the data set.

3. Uses of Digital Cadavers

The UW-L project is adopting a model where students will over the duration of a semester augment traditional lecture/laboratory instruction with periodic assignments involving the digital cadaver. Digital cadavers will be used in much the same manner as physical cadavers. Students would be given assignments involving specific anatomical structures. For instance they might be asked to demonstrate the relationships between the trachea, carina, bronchi, esophagus and aorta. Imaging would be used to directly support the objective of portraying anatomical features within the visual context of surrounding structures. Students will collect the results of their efforts into digital notebooks of images, annotations, and commentary. These notebooks would form an objective basis for grading their work.

In the fall of 1995 the use of digital cadavers will begin to be integrated into instruction in anatomy and physiology courses offered through the Department of Biology and serving a variety of allied health programs. Due to limited equipment availability, initial use will be restricted to a few sections of the class. In addition to student and faculty use we also foresee the imaging capabilities of the system being employed in an increasing array of distance learning and extension offerings being developed by the university.

4. Software Development

Since our primary objective is to give undergraduate anatomy students direct control of the tools for imaging the VHD data, the tools must assume little or no understanding of the data and imaging methods used. Students and instructors are concerned with images which reveal the detail of anatomical features not with how the images are computed. Consequently, an explicit design goal of the software tools is that they require as little input from the user as possible. This implies extracting as much information and heuristic hints as possible out of the input provided. If the user wants to generate a particular cross section, it should be possible to specify the cross section by pointing at landmarks within the VHD section images. If the user wants to generate a volume rendered image, image parameters such as lighting and opacity should be automatically computed in such a way as to give standard predictable results. Such a design may limit the fine control of images produced by the system, but we do not believe that it will in anyway detract from the utility of the images generated with respect to revealing anatomic structure. It is certainly the case that whatever limitations are introduced are more than compensated by gains in ease of use. We believe that this trade off is entirely appropriate for our intended users.

The software is being developed according to a client/server model. Imaging computations are performed on a server hosting the data set, while students interact with client tools. The software consists of three logical groups.

This organization keeps data intensive operations close to the data source.

4.1 Data Preparation and Reduction

The VHD cryosection images consist of a 2048x1216 raster of 24 bit RGB pixels making each section approximately 7MB. Images of this size are awkward to use at best. Loading and displaying an image takes a significant time and the resolution exceeds the screen resolution of almost all displays. We reduced the image size by cropping the extraneous border from each image. This reduced the image size to approximately 5MB. We color edited the images to remove the blue matrix in which the cadaver is embedded and with which cavities are filled. We also found it necessary to shift groups of image sections along their width in order to obtain correct registration of the sections down the length of the cadaver.

To further reduce the image size we computed a composite color histogram of the section images. The color histogram showed that pixel colors were highly organized into a planer subspace of the RGB volume extending out from the origin just below the main diagonal of the RGB volume and angling towards the red axis. Using this color histogram we choose a palette of 256 colors concentrated in the region occupied by image pixel colors. Each image was again color edited using this palette, scaled by one half and used to produce an 8-bit paletted TIFF image. This reduced the image size to approximately 0.4MB. The resulting TIFF images provide good color fidelity, distinguishing small gradations in color and texture. The TIFF images may prove to be of sufficient quality that we will not need to maintain the full RGB data on line. Either the full RGB or palletized TIFF images can be used for viewing individual slices of the data.

4.2 Cross Sectional Images

The VHD cryosection images are axial sections. Common practice dictates that certain parts of the anatomy be portrayed by sections parallel to specific planes. This is particularly true of the head where sagittal sections are commonly used for visualizing specific structures. Nonorthogonal sections are appropriate where it is desirable to construct a section cutting through specific landmark structures or when it is necessary to adjust for deviation of the body with respect to the main axis of the VHD volume. The browser interface interprets lines drawn on a VHD slice as requests for the construction of a cross section image orthogonal to the VHD slice and with a horizontal axis and extent determined by the line. If in addition to a line in a VHD section, the user specifies a landmark point contained within another VHD section, the system will construct a nonorthogonal section containing the line and point.

4.3 Volume Rendered Images

Volumetric imaging is a technique for visualizing the three dimensional structure of a volume of scalar or vector data [2,3]. In medical imaging the technique is typically applied to CT and MRI data, but in other fields the technique has been employed with a variety of data types. Since the VHD cryosection images collectively form a three dimensional volume, with each voxel representing tissue color, the VHD cryosection images as well as the radiological data are suitable for volume rendering.

Volumetric imaging is recognized as having a number of advantages over other techniques such as isosurface rendering. In isosurface rendering geometric surfaces are extracted from the data, usually in the form of triangle meshes, and rendered as solids which with appropriate hardware can be manipulated in real time. The surfaces approximate locations within the data of equal value (e.g. density, color). In order to construct the surface it is necessary to choose threshold values for the surfaces. This choice can sometimes obscure features of the data.

While there are several approaches to volume rendering (e.g. ray casting, voxel splatting) the basic technique is to accumulate, for each pixel in an image plane, the color and opacity of voxels lying along a ray emanating from the viewers position and passing through the image plane pixel. The consequence of this approach is that every voxel of the data has the potential to contribute to the rendered image depending upon its color and opacity. As with ray tracing algorithms it is possible to incorporate sophisticated lighting models. The high computational requirements of volume rendering algorithms in general exclude the possibility of interactive frame rates unless significant constraints are placed on the data set size and image sophistication. Consequently, volume rendering systems are generally used to produce individual images or successive frames of a movie. The problem is purely one of computational speed however, and every year the scope of what can be interactively rendered increases. The main advantage of volumetric imaging is it ability to visualize the entire data set without the necessity of extracting geometric surfaces. These essential characteristics of volume rendering provide excellent support for our objective of providing students with tools that will enable them find and exhibit anatomical detail within the visual context of surrounding structure.

Our browser interface provides access to a volume rendering system running on the compute server. The volume rendering system implements a ray casting algorithm which produces perspective images of the data volume. The algorithm assumes a large data set. It can operate on both the cryosection and radiological data. In order to reduce the size of the cryosection data the system operates on the half resolution TIFF raster data. This presents the data as a volume of single bytes, each byte encoding a color from the 8 bit TIFF palette. A false color map for the radiological data and a correspondence between opacity and tissue type can be provided. The system incorporates a modest lighting model and attempts to compute smooth gradients for use with lighting.

The main obstacle to using volume rendering systems is the number of parameters involved in specifying an image. It is usually necessary to have at least a minimal understanding of how the image is computed in order to efficiently use the system. For the students and faculty who will be using our system, computational issues are an obstacle to use. Their only concern is the anatomical structure and its context. In order to satisfy this need we are designing our system so that all parameters not having to do with locating the structures to be rendered, will be determined for the user in such a way as to produce reasonable results. At a minimum the user indicates a focus point within a VHD section. Using a virtual trackball interface, the viewing direction with respect to the section is indicated. If in addition to a focus point, the user specifies a bounding box within a VHD section, an appropriate viewing distance can be computed. Lighting characteristics and location are determined relative to the viewing position. The system will produce either individual images, stereo pairs or movies.

4.4 Digital Staining

Frequently, opacity is the most difficult parameter to control in volume rendering. An interface consisting of a system of sliders for controlling opacity of different selections provides no usable information for our intended users and without interactive feedback becomes little more than a trial and error process. From the users point of view, who thinks in terms of the desired image and not in terms of opacity values, the specific value of an opacity parameter is meaningless. From the user's point of view it should be sufficient to indicate that one or more structures are of primary interest and should be opaque while other structures are for context and should be rendered partially transparent with all remaining structures fully transparent. From this information the level of opacity required to keep the primary structures visible in the final image can be approximated. This can be achieved by following a cluster of test rays back from points within the primary structure to the view point, accumulating information about the intervening material. In order to carry out this computation, the user must have a convenient way of indicating the extent of the structures to be visualized.

To address this problem we are developing an interface for "digital staining". As the name implies, "digital staining" allows a user to identify and mark anatomical structures. Staining is not a drawing or outlining process. It is a way of using user gestures to direct a breadth first, limited, pixel filling process. The process combines information extracted from user input with image information such as color differences and gradient magnitudes. Pixels are filled based upon their distance in RGB space from a reference sample of colors, their distance from a user gesture position and their continuity with previously filled pixels. The feel of the interface is intended to give the impression that the stain flows out from the users pointing position and can be pushed about as desired.

Staining does not alter the original data but operates as an overlay to the VHD sections. Staining information is stored separately in a compact run length encoded form. Users are able to treat staining information as a document type, opening and saving as desired. The viewer interface allows multiple stain documents to be in use at once. When passed to the volume renderer, the user chooses which stains are to be used and how they are to be interpreted. It is also possible to create geometric stains (cylinders, slabs etc.) to be used by the volume renderer in creating cutaway effects.

Staining does require work but visually isolating structures within a VHD section also requires effort. Our goal is to make the staining process no more of a burden than visually identifying structures within a VHD section.

4.5 Notebooks

The idea of an anatomical notebook is central to this project. Along with examination of the VHD sections students should record their observations just as they would keep a laboratory notebook in other sciences. Where a student might, in other circumstances, have kept notes and sketches about dissected anatomy, students will now collect and annotate images derived from the VHD.

The communication between data browser and notebook is bidirectional. Images generated using the browser can be dragged into the notebook. Since image files contain a record of the parameters used in their generation, dragging an image from the notebook to the browser allows the browser to extract the parameters for reuse or modification.

Currently, we are using the standard RTFD editor supplied with NEXTSTEP for the notebook interface. The editor has good support for embedded images and objects of other kinds, including other document types such as stain overlays. Objects can be dragged between a notebook document and other applications. It also incorporates a hypertext link facility. Depending on our future needs we will investigate other options including WYSIWIG HTML editors.

4.6 Current State of Development and Future Directions

We have developed initial versions of the browser interface and imaging servers. In addition to making continual refinements to the user interface, in response to feedback from students and faculty, we are particularly interested in refining the staining interface and are looking at the use of pressure sensitive pen tablets in place of mice in order to extract more information out of user gestures. We are also investigating ways of improving the rendering algorithms by distributing the computational budget based upon image characteristics such as viewer distance and data volume size.

In the future we hope to provide a minimally interactive volume rendering system based upon the single buffer, random pixel update scheme outlined in [1]. The idea is to approximate an image while the user is rotating the data. The image would be produced by continuously recomputing individual pixels of the image, updating the graphics frame in single buffer mode. Pixels chosen for updating would be randomly chosen from the image frame with a bias for pixels and neighboring pixels where the image exhibits significant change between updates. When the motion is stopped the process would continue to update pixels chosen at random until the entire scene was recomputed.

5. Facilities

The client tool interfaces to the VHD are being developed and run on 100MHz Pentium PCs running NEXTSTEP. The imaging daemons and data are being served by a 4 processor HP 9000/K200 running HP-UX 10.01 with 256MB of memory and 24GB of disk space. In addition to this equipment we have several SGI Indigo2 workstations with stereo glasses available for polygonal rendering.

6. Conclusion

The UW-L project is exploring innovative ways of using anatomical data such as that provided by the VHD to provide undergraduates in anatomy and physiology with learning experiences otherwise unavailable to them. Individual work with physical cadavers reinforces a number of physical and mental skills. Most obvious is the physical skill of dissection. Less obvious, but equally important is the mental skill of using existing knowledge to reveal and present new anatomical features. This is the process by which knowledge of human anatomy is acquired and reinforced through repeated observation and use. By providing students with "personal digital cadavers" the UW-L project is developing a learning environment which rewards the effort expended in isolating and displaying anatomical structures. In this regard the critical thinking skills needed to work effectively with digital cadavers closely parallel the skills required for working with actual cadavers. The software environment being developed supports intuitive access to sophisticated image rendering tools as well as the ability to maintain a written, visual record of all work performed. Using this environment students are encouraged to take responsibility for their anatomical studies, building upon existing knowledge as they acquire new information.

7. Images



The basic browser interface and constructed cross section.


The optic nerves and eye rectus muscles rendered under a semi-transparent boundary cutaway plane. (cross-eyed stereo pair)


An isolated cervical vertebra with disks, spinal cord and nerve bundles leaving vertebra. The vertebra is rendered partially transparent. (cross-eyed stereo pair)


A cervical vertebra embedded within quarter section of neck and head. (cross-eyed stereo pair)


Opaque section of shoulder with scapula and clavicle. Muscles rendered as a semi-transparent boundary. (cross-eyed stereo pair)

8. References

[1] Bishop, Gary, Henry Fuchs, Leonard McMillan and Ellen J. Scher Zaiger. Frameless Rendering: Double Buffering Considered Harmful. Proceedings of SIGGRAPH '94. In Computer Graphics Proceedings, Annual Conference Series, 1994, ACM SIGGRAPH, pp. 175-176.

[2] Elvins, T.T., "A Survey of Algorithms for Volume Visualization", Computer Graphics, Volume 26, Number 3, August 1992, 194-201.

[3] Elvins, T.T. and Nadeau, D.R., "NetV: An Experimental Network-based Volume Visualization System," Proceedings of the IEEE Visualization '91 Conference, IEEE Computer Society Press, October 1991, 239-245.

[4] National Library of Medicine (U.S.) Board of Regents. Electronic imaging: Report of the Board of Regents. U.S. Department of Health and Human Services, Public Health Service, National Institutes of Health, 1990. NIH Publication 90-2197.