The Parametric Human Project RSS Feed

Preview Our Data! (Blog Entry)

Thu, 12 Aug 2010 15:05:59 -0700

We are very excited to announce that the preview of our bone database is now online!

As you know, we have been scanning human bone samples for the past few months. We are now working on the infrastructure to make this data available for download. However, we wanted to give you a taste of what we have achieved. Currently, we have two mostly complete bodies scanned (awkwardly named "box171" and "SkeletonBox19") and we are beginning work on the third one! There is also a set of about eight isolated tibia bones.

Click here or on the "Data Sets" link above to go there now!

As you can see, we've included with each scan the number of points comprising it and a bounding box. On average, most of our models have at least one million points. In fact, here are some statistics:

	Number of scans	Total point samples
box171	174	449,130,596
SkeletonBox19	176	416,087,933
SkeletonJ-19	10 so far	9,645,398 so far
Tibia samples	8 so far	56,487,265 so far

As well as working on the download framework, we are verifying all the data, selecting a consistent, more correct naming scheme, and generating surface meshes at various scales for further processing.

Please send us any comments, requests, or ideas. Just click here to tell us what you think!

Interactive watershed code! (Blog Entry)

Tue, 03 Aug 2010 12:59:30 -0700

It's been a long time coming, but it's finally here: the code for the Interactive Watershed Transform we used to fully segment and separate individual bones from CT scans!

As shown in our previous post, this code is built as a module for MeVisLab. It was originally developed for MeVisLab 1.6.1 using Microsoft Visual Studio 2005, but can easily be ported to MeVisLab 2.0 and MSVC 2008. It has not been tested with MeVisLab 2.1.

If you download and use this code, please let us know how you are using it and how we can change/improve it! Just click on the "Contact us" link above to find where to reach us!

Click here to download and enjoy!

How Human Bones Grow (Blog Entry)

Thu, 15 Oct 2009 14:16:40 -0700

Most people I speak to are surprised to find out that the bones in the arms and legs are separated into three pieces in children and only fuse together late into the teen years. However, this well documented fact helps pediatricians determine the biological maturity (age) of a child by analyzing the X-ray of the child's hand and wrist (see, for example, the Hand Bone Age book). Further, a company by the name of BoneXpert has created an automated system to accomplish this task! What follows is a brief description of how human bones grow.

X-ray photo of 7 year old boy. Note how the knuckles are separate from the rest of the hand bones (metacarpals) and how the wrist bones near the thumb are just beginning to form.

I am sure by this point many of you are wondering, especially given the image above, what is holding the body together in a child if there is no bone? The answer is cartilage, a tough elastic tissue that serves as a model for the shape of the bone.

Before we discuss how bones are built, let us take this opportunity to describe the types of bones and their structure. Human bones are generally classified as long bones, flat bones, and irregular bones. Long bones include the bones of the arms, legs, hands, and feet. Flat bones are found in the cranium surrounding the brain. Irregular bones are those which cannot be described by either of these terms. They include the vertebrae and the bones of the wrists and ankles.

Regardless of the bone type, all bones are arranged in two types of structures visible to the unaided eye (macroscopically). Cortical or compact bone is found mainly in the shafts of long bones. It is very dense bone, lacking visible spaces to the unaided eye. It's function is mostly mechanical, giving the shaft of the bone its strength. The other kind of bone is spongy, trabecular, or cancellous bone, which is organized as a network of branching fibers called trabeculae. The images below show examples of both types of structures.


Cross section of a human radius bone showing the structure of cortical bone.	Proximal end of a human radius with damage to the outer layer, revealing the spongy bone beneath.

CT slice of femur along the shaft. Cortical bone generates a strong response in the image.	CT slice of femur near the knee (note the patella above). Spongy bone does not appear as bright in the CT image due to its reduced mineral density.

Microscopically, there are also two types of bones. Woven bone is a mechanically weak arrangement characterized by the random organization of fibres. It is very atypical, appearing in the beginning of bone development and in fracture repair. The more common kind is called lamellar bone, which is very strong due to its highly-organized, parallel layers (called lamella) of minerals.

Cortical bone has a special arrangement of lamella called osteons. In an osteon, the lamella are organized in a cylindrical fashion around a canal which carries blood vessels and nerves (the Haversian canal). Some of the osteoblasts become trapped in small crevices (lacunae) between the lamella and are then known as osteocytes. The osteocytes between each lamellar layer are in contact with each other through tiny canals (canicullae) which allows them to exchange nutrients within the otherwise stiff mineral structure of the bone. This structure is not present in the bony trabeculae since osteocytes receive their nutrients from the bone marrow.

The process of bone generation is called ossification. There are actually two types of ossification:

endochondral ossification:: where the bone replaces the cartilage model. Occurs mostly in the long bones.
intramembraneous ossification:: where there is no cartilage model. Occurs mostly in flat bones.

Let us describe intramembraneous ossification first, as it is a simpler process. Intramembraneous ossification occurs when a number of cells in the developing body differentiate into osteoblasts, or bone depositing cells. These cells secrete a compound called osteoid which later becomes calcified and results in the development of bony trabecula.

Endochondral ossification begins in the middle of the cartilagineous model of the bone shaft, also known as the diaphysis. Here, the intercellular matrix becomes calcified and prevents the diffusion of nutrients to the cells within the bone. As these cells die away, they leave a network of calcified cartilage behind. At this point, blood vessels arising from the tissue surrounding the cartilaginous bone model (the periousteum) penetrate the calcified cartilage. Osteoblasts use this calcified cartilage as a scaffold to begin laying down osteoid, forming trabecula. Osteoblasts also appear in the periousteum and begin to deposit bone against the shaft of the bony cartilage, producing a bone collar. This deposition results in a lamellar arrangement called circumferential system, which surrounds the outer surface of the bone (the missing outer layer in the image of the spongy bone of the radius above). Another kind of cell, the osteoclasts, appears inside the shaft and begins to break down the spongy bone from the inside, forming the medullary cavity in which the bone marrow resides.

Here, the bone begins to grow in two directions. It increases in thickness by laying down bone from the periousteum (appositional growth) and it increases in length as the osteoclasts break down the spongy bone toward the ends of the bone (the metaphysis). This entire region of growth is known as the primary ossification center.

At birth or even later, secondary ossification centers appear at the ends of the long bones, laying down spongy bone. These bony end pieces, the epiphyses, are separated from the metaphysis by a layer of cartilage called the epiphyseal plate. As the cartilage in the metaphysis is turned into bone, more cartilage is generated, thus lengthening the bone. At some point, near 20 years of age, this cartilage is no longer regenerated, and the cartilagenous epiphysial plate disappears. The diaphysis and epiphysis fuse together achieving the final shape of the bone.

However, changes in the bone structure continue throughout our lifetime. Osteoblasts and osteoclasts are continuously remodelling the bony structures, replacing old minerals and cells with new ones. New osteons are built whenever there is enough space for them to appear, even if the old osteon is not fully removed. This produces the third kind of lamellar arrangement within the bone called the interstitial system. Osteoporosis results when more mineral is removed by the osteoclasts than what is deposited by the osteoblasts.

The images below show the epiphyseal line, the line at which the diaphysis and epiphysis of the bones fuse.


Distal epiphysial line of the femur.	Matching humerus bones, with one of the bones separated at the epiphysial line.

Another Data Source (Blog Entry)

Wed, 30 Sep 2009 15:56:28 -0700

So far, we have discussed using medical images to extract human bone shapes for analysis. This is not, however, the only source of data for human bone shapes. Another accurate source of human bone geometry are human bones from human remains.

Actual samples of human bones are quite common since, of all the organs in the body, bones are the most durable. Skeletons can be recovered many years after the individual's death, as long as the body has not been cremated. They are also easy to handle and store and do not require as much care as other organs to maintain. For this reason, many human bone collections exists throughout the world for studies in anthropology, forensics, and medicine. The various collections are listed at this website.

Note that bony remains are very different that the actual living bone. Human bone is composed of living and non-living components. Living components include bone-depositing cells (osteoblasts), bone-resorbing (removing) cells (osteoclasts), blood vessels, and blood cells to name a few. The non-living tissue is a mixture collagen and hydroxyapatite and provides the structure and strength of the bone. This is the bony tissue that remains long after the living tissue has died.

The image above shows 5 human tibia samples that have been cleaned for use in teaching. Preparing these samples is a labor intensive process. First, all the muscle, tendon, and ligament tissues must be removed from the bone. This is done through boiling, composting, or using dermestid beetles. Bones also store fat inside (yellow marrow), so they must be degreased. Finally, they are typically whitened using hydrogen peroxide before they are sold for academic use.

To create a digital model of the shape of the bone we use a FARO Laser ScanArm mounted on a FaroArm Quantum measuring arm. This system is simply a hand-held, 3D laser triangulation scanner that generates 3D position measurements on the surface of an object quickly and accurately. The principles of how it works is very simple: the scanner shines a laser beam, spread out in a thin line, on the object to be scanned. Then, a camera placed at a specific angle to the beam captures the image and looks for the bright red laser light. Using the angle between the beam and the camera and the position of the red beam in the camera image, a 3D point is computed where the surface of the object reflects the laser beam. A wonderful depiction of this is shown here. The images below show the scanner in action and the results on the computer screen.

We have set up a lab to scan complete human skeletons such as the sample shown below. We are very near completion of our first, mostly complete skeleton (unfortunately, we are missing the skull, the sacrum, and one tibia).

Our scanning sessions proceed as follows: in the morning, we check our e-mail or read the latest news while we sip our coffee and allow the laser scanner to warm up. Next, we recalibrate the scanner as seen below, on the left. Finally, we grab the next bone in the spreadsheet and begin to scan. Our spreadsheet is a template for a database we will design and build which will include, in addition to all of the bone shapes, the geographic origin of the specimen, sex, and age at death (if known), and tagged landmarks of salient anatomical features. We are also including some extra slots for anomalies--the specimen we are currently scanning, for example, has an extra pair of cervical ribs! This will require massive amounts of storage: our scan files run between 1 and 8 million vertices per bone!

We typically scan each bone twice, once from either side (for example, a superior and inferior view for a vertebra, or palmar and dorsal views for a metacarpal). There are, however, some bones which require more than two scans to capture all the nooks and crannies in their shape digitally (the triquetral, for example). Due to the scanning technology, some of the scans will invariably have holes where the laser light cannot reach. So far, this has only been the case for some of the cervical and thoracic vertebrae, but we expect to have the same issue with the skull. The scanning software provides features that allow us to clean up the scans and register the pieces together into the final product.

Using the Interactive Watershed Transform (Blog Entry)

Tue, 22 Sep 2009 17:50:07 -0700

Given all the previous research work to separate bones in CT scans, we decided to give some of the techniques a whirl. We showed in our previous post our results using the c-plane and sheetness techniques. Here we discuss our implementation of the Interactive Watershed Transform (IWT).

This algorithm, based on the watershed segmentation algorithm, was developed by Hahn et al. and is very well described in Hahn's Ph.D. thesis, available here. The idea behind the watershed segmentation is to consider the gray levels of the image as a height field, and imagine placing a drop of water onto each pixel. The water will follow the terrain slope until it reaches the bottom of a valley. The segmentation is done by grouping pixels whose water drop end up in the same basin--that is, those that belong in the same watershed.

The idea just described of following the terrain slope is one method of implementing the watershed algorithm. A second way is to "flood" the image terrain with water starting at the lowest point. As the water rises, new basins will begin to flood and watersheds will merge. The user can control at what point to stop the flooding and thus where to end the segmentation.

One of the main drawbacks of the watershed segmentation algorithm is that it tends to over segment the image, creating many small basins that do not correspond to image features. A solution for this is to pre-flood the image in order to merge some of these smaller basins together, especially those which are separated by a low-height watershed line. Also, given a segmented image, the user can request certain basins to be merged, thus obtaining the segmentation he or she wants.

The insight behind the Interactive Watershed Transform is to cache the basin creation and basin merging events of the flooding algorithm so that the entire segmentation can be replayed interactively. Combined with pre-flooding and user selection of associated basins, this becomes a powerful segmentation pipeline. It is not a fully automated procedure, but it lets the computer do the hard computation while the human user interactively guides the segmentation process to obtain the results he or she needs.

Hahn describes a procedure for using the IWT to separate the carpal bones. First, the image is inverted so that the brightest pixels, corresponding to the bone, become the darkest, thus becoming the basins for the watershed segmentation. The user then uses an interactive IWT to select the basins corresponding to the bone he or she wants. This segmentation does not lead to the accurate bone boundaries that are obtained via thresholding. In the final step, Hahn combines the thresholded image and the IWT segmented image using a logical AND operation to obtain the final result.

We decided to test the IWT segmentation algorithm to separate the ankle bones of the Visible Female dataset. As can be seen in the images above, the talus (ankle) bone is attached to the tibia (shin) bone on the top (superiorly), the calcaneus (heel) bone in the rear (posteriorly), and the navicular bone to the front (anteriorly).

We implemented the IWT procedure using MeVisLab, a very versatile program for processing and visualizing medical images. It combines the versatility of the Insight Toolkit (ITK) and the Visualization Toolkit (VTK) with the scene management capabilities of Open Inventor. MeVisLab also provides its own image processing pipeline, is easily expandable by writing custom nodes, and is scriptable. The codebase is actively developed and a slightly restricted version of the program is available for non-commercial use.

We wrote a number of custom nodes to do the job. The basic pipeline is shown in the image above. The IWT algorithm was encapsulated into two nodes (with really long, self-descriptive names). The IWTProcessor computes all the basin creation and merging records, while the IWTBuilder takes as further input the basin markers provided by the user interactively. The output from the IWTBuilder is combined with the thresholded input image (at the "Bone Selector") before it is sent on to the output.

This entire network was then encapsulated in a single macro node (a node with an inner node network) so that a custom user interface could be built. Among the controls provided by the interface is a selector for the currently viewed slice, the pre-flooding height, and the ability to place markers in an image. There are four views of the image in the user interface. From left to right, top to bottom, they are:

the original image,
a visualization of the watershed basins,
the thresholded image, and
the final output image.

The interface in action is shown in the video below. The IWT algorithm is only applied to a small region of interest surrounding the talus bone. The user first searches for a slice where the talus bone is clearly discernible, the increases the pre-flooding value until most of the image basins have merged. Then, he places two markers in the basin image. This is enough to fully segment the bone, as he later verifies by looking slowly through all the slices.

(To view this video you will need to enable JavaScript or visit our website.)

The animated images below show the original ankle bones and the segmented talus bone. Note how the talus surface is closed at the articular interfaces. One drawback of this system is that the final image is dependent on thresholding, so it is possible for the final result to have gaps where the bone is thin--this is visible in the talus bone below. Another issue we encountered was the merging of basins corresponding to separate bones, even with the pre-flooding set very low. In such cases, it was impossible to get an accurate segmentation. Other than that, we found the technique very easy to use and provided great results!

(To view this video you will need to enable JavaScript or visit our website.)

More About Bone Segmentation from CT Images (Blog Entry)

Thu, 17 Sep 2009 14:29:44 -0700

Thresholding the CT image is a simple way to separate bone tissue from soft tissue. Some of the drawbacks of this fully automated technique, however, are the merging together of nearby bones and the creation of holes in the data where the bone is thin or very spongy. In what follows, we discuss some of the segmentation techniques developed to improve the correct segmentation of bony structures from CT scans.

First, let us consider a simple extension to the basic thresholding algorithm. What if we could vary the segmentation threshold as we moved along the image, based on a property of the bone? For example, the skull has many bony regions where the bone is paper thin and flat. Westin et al. developed a technique that lowers the segmentation threshold adaptively over the image depending on how close the pixel's neighborhood resembles a plane.[2]

The idea is very simple. First, Westin generated a 3x3 tensor representation of the orientation of the image signal using 6 oriented filters. By analyzing the eigenvalues of this tensor, a measure c-plane describing how similar the signal (CT image response) in the region near the pixel resembles a plane (for the mathematically inclined, we will describe the mechanics of this a little further below). Finally, during segmentation, the threshold at each pixel is reduced by the c-plane measure times a constant, thus allowing more thin, plane-like structures to show up in the final image.

Westin et al. later extended this work to enhance the bony structures of interest in the original image to make segmentation easier[3]. This time, instead of using an adaptive thresholding scheme, they compute a new set of adaptive filters based on the orientation tensor. This new filter bank consists of a single low pass filter and six adaptive, oriented filters. The oriented filters are constructed so that their effect is cancelled in areas where there is no definite orientation (noise). After enhancing the data with this filter bank, the image is thresholded and labeled into connected components. Finally, empty voxels contained inside closed structures were merged with the surrounding structure.

Descoteaux et al. also developed another measure for "planeness" or "sheetness." [6] Their measure is also based on the analysis of the eigenvalues of 3x3 tensors defined on the image signal; however, the tensors they analyze at each pixel are Hessian matrix and the outer product of the image gradient. The insight here, and in the work above, is that only one of the eigenvalues will have an absolute value much greater than zero when the object is sheet-like. Descoteaux notes that if two eigenvalues have magnitude much greater than zero, then the region is tube-like, and if all three are much greater than zero, the region is blob-like. If all three eigenvalues are near zero, then we have found a noisy background pixel. Descoteaux's sheetness measure which contains a sheet enhancement term and blob and noise suppresion terms.


Standard thresholding	Using c-plane	Using sheetness

The images above show the results we obtained using these adaptive thresholding techniques using the eigenvalues of the Hessian matrix computed over the head of the Visible Woman. Note that although there is some enhancement of the bones in the nasal cavity, the orbit, and the cheekbone, the change in quality is not significant, and some spurious noise is now also visible.

Thresholding and adaptive thresholding are not the only algorithms to have been applied to the bone segmentation problem. Tagare et al. used a path optimization algorithm requiring user initialization to segment individual slices of the carpal bones. Their method used a dynamic programming algorithm to compute an optimal, eliptically parametrized path taking into account the image value, it's gradient, and the angle of the gradient with respect to the X axis [1].

Descoteaux et al. used their sheetness measure to create a vector field which guided a developing contour surface toward the surface of the bone using geometric flow [6].

Kang et al. used a multiple step procedure for segmentation [5]. First, they used a region-growing algorithm which combined properties of the global and local histogram around a pixel. Next, the interior of the bone volume was labeled using a combination of morphological closing operations and filling of closed contours. Next, they find the internal surface of the cortical region of the bone by computing a 50% relative threshold within the bone volume along the surface normal, using a running average to smooth out the noise. In the fourth step, a human operator has the opporunity to correct any of the previous steps. For bones that are close in proximity, a mask is required to isolate the region of interest to avoid connected bones.

Using a combination of techniques, Sebastian et al. presented some promising segmentation results in 2D [4]. Their idea is to combine the deformable models of curve evolution with region growing and region competition. In their skeletally-coupled deformable model, the inter-region skeleton, which is the predicted locus of collisions between two growing regions, is used to couple together the movement of the regions before they collide and prevents them from growing too fast or too slow with respect to each other.

Recently, Scherf and Tilgner have introduced a new segmentation algorithm optimized for analyzing trabecular bone from micro-CT scans in the context of physical anthropology [8]. They first compute Canny edges of the image. In the novel step of their algorithm, they cast rays along the surface normal of each contour point until it intersects another contour, thus labeling the non-bony regions of the image. Finally, a connected component algorithm selects the non-highlighted region as bone.

Not all researchers have aimed at developing fully automated algorithms. In fact, it is best if the computer does 99% of the work and a human user verifies and corrects the result. That is the approach of the final two algorithms we discuss today.

Liu et al. designed an interactive system for bone segmentation based on the graph cut algorithm [7]. After applying an edge-dectector to the CT image, the user applies seed points in any location of the 3D volume. The seed point label is expanded to all nodes in the graph less than a user-selected distance from the seed. Also, the edges are weighted in such a way that the graph cut algorithm attempts to reduce the number of edges in the cut, under the insight that there are few connections between separate bones.

Finally, Hahn describes an interactive watershed transform that is useful for segmenting the carpal bones and separating bone from white matter in an MRI [9]. His idea is to cache all the basin creation and merging events into a simple data structure as the watershed algorithm progresses. Later, the user interactively modifies a pre-flooding constant (that helps with the over-segmentation problem) and places label points in the image. The watershed labels created in this way are used to separate the merged bones segmented using standard thresholding.

References

[1]	H. D. Tagare, K. W. Elder, D. M. Stoner, R. M. Patterson, C. L. Nicodemus, S. F. Viegas, and G. R. Hillman. Location and geometric description of carpal bones in ct images. Annals of Biomedical Engineering, 21:715-726, 1993.
[2]	Carl-Fredrik Westin, Abhir Bhalerao, Ron Kikinis, and Hans Knutsson. Using local 3d structure for segmentation of bone from computer tomography images. Computer Vision and Pattern Recognition, IEEE Computer Society Conference on, 0:794, 1997. [ DOI ]
[3]	Carl-Fredrik Westin, Simon K. Warfield, Abhir Bhalerao, L. Mui, Jens A. Richolt, and Ron Kikinis. Tensor controlled local structure enhancement of ct images for bone segmentation. In MICCAI '98: Proceedings of the First International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 1205-1212, London, UK, 1998. Springer-Verlag.
[4]	Thomas B. Sebastian, Huseyin Tek, Joseph J. Crisco, Scott W. Wolfe, and Benjamin B. Kimia. Segmentation of carpal bones from 3d ct images using skeletally coupled deformable models. In MICCAI '98: Proceedings of the First International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 1184-1194, London, UK, 1998. Springer-Verlag.
[5]	Yan Kang, Klaus Engelke, and Willi A. Kalender. A new accurate and precise 3-d segmentation method for skeletal structures in volumetric ct data. IEEE Transactions on Medical Imaging, 22(5), May 2003.
[6]	Maxime Descoteaux, Michel Audette, Kiyoyuki Chinzei, and Kaleem Siddiqi. Bone enhancement filtering: Application to sinus bone. In in Proc. Int. Conf. Med. Image Comput. Comput. Assisted Intervention, 2005, pages 9-16, September 2005.
[7]	Lu Liu, David Raber, David Nopachai, Paul Commean, David Sinacore, Fred Prior, Robert Pless, and Tao Ju. Interactive separation of segmented bones in ct volumes using graph cut. In MICCAI '08: Proceedings of the 11th international conference on Medical Image Computing and Computer-Assisted Intervention - Part I, pages 296-304, Berlin, Heidelberg, 2008. Springer-Verlag. [ DOI ]
[8]	Heike Scherf and Rico Tilgner. A new high-resolution computed tomography (ct) segmentation method for trabecular bone architectural analysis. American Journal of Physical Anthropology, 140(1):39-51, March 2009.
[9]	H. K. Hahn. Morphological Volumetry: Theory, Concepts, and Application to Quantitative Medical Imaging. PhD thesis, University of Bremen, January 2005.

This section was generated by bibtex2html 1.92.

Thresholding: A Basic Segmentation Technique (Blog Entry)

Wed, 09 Sep 2009 14:19:05 -0700

The 3D data obtained from a CT scan provides an accurate snapshot of a human body part. Additionally, it is well suited for imaging bone data, since bones appear as the lightest areas of the scan and in high contrast (see the previous post for more details).

In order to work with the bone data directly, we must separate it from the rest of the tissue data in the scan. We can easily achieve this segmentation by carefully labeling all bright points in the scan as bone and all darker points as air. Thus, we would separate all of the bone voxels from the tissue voxels.

An easy algorithm for doing this automatically is the thresholding segmentation algorithm. This algorithm sets the voxels with a value higher than a given threshold to white and sets all other voxels to black. This works well for bone since its output response to the CT X-ray is higher than the soft tissue surrounding it.

On the left is an original slice from a CT scan of the ankle from the Visible Human data set. In the middle is the same image after applying the thresholding algorithm. On the right is a visualization of the entire ankle volume after thresholding. Note how the muscle and other tissue is no longer visible.

Although this procedure is very easy to implement, fast, and fairly accurate, we cannot directly use the resulting voxel data to anaylize the shape of the bones. To use computer graphics techniques, it is best to convert the data to a smaller and more flexible surface representation such as a triangle mesh. Using a triangle mesh, we can easily represent the exterior and interior surfaces of the bone and efficiently apply geometric analysis procedures.

In order to extract a triangle mesh from the bone data, we employ an algorithm similar to thresholding but outputing surface data. The most common algorithm in this case is marching cubes. This algorithm walks across voxels comparing their image values. Whenever it detects a crossing of the threshold value, it knows that the surface crosses through that voxel. By analysing the crossing configuration it then places the triangles in the voxel and generates a surface. The next figure shows the output of a variant of marching cubes when applied to the ankle data shown above.

The marching cubes algorithm is very fast and produces great results. For our purposes, however, it has some drawbacks. First, as can be seen in the image above, the bones are extracted as a single surface and the boundaries between the bones are molded together. Second, as can be seen in the images below, the places where the bone is very thin or very spongy do not produce a signal strong enough to surpass the requested threshold. This leads to holes in the bones as highlighted by the arrows in the images below.

The thresholding parameter can be changed. However, our experience has shown that some areas of the body have bones that are too thin to be extracted correctly. In future posts, we will discuss some of the research work that attempts to solve this problem.

In the meantime, if you are interested in trying out these techniques for yourself, download the Visualization Toolkit (VTK) and give it a shot. This article is a great place to try out segmenting the skin and bones of the Visible Human data set.

Scanning Technologies (Blog Entry)

Thu, 27 Aug 2009 15:09:30 -0700

Our goal is to build a statistical model of shape variation in the human skeleton due to age, sex, and geographic origin. We want to compare and quantify differences in shape among a large collection of human bone shapes with enough digital specimens that we can create a model of how the bones' shape varies using statistical techniques. Note that we are interested in variations of the complete 3D shape and not simply landmark measurements.

The first challenge in accomplishing this task is obtaining a large database of bone shapes of people from different age groups, sex, and geographic origins.

Two dimensional X-ray images are used by doctors and radiologists to visualize fractured bones, tumors, and blood vessels. Bones, in particular, have a high absorbency rate for X-rays, producing a lighter color in the final image. This contrast enhancement is what makes X-rays particularly useful for imaging bones.

A 3D version of the X-ray image is the computed tomography scan, or CT scan for short. CT scans can be thought of as a sequence of X-ray images focusing on a small, planar section of the subject's body. By combining a large number of these sectional images, a 3D volume image of X-ray data is built. CT scans are fantastic for visualizing bone shapes, since they produce 3D images in which the bones show up in high contrast.

The images above show 3 slices from a CT scan of the ankle from the Visible Human data set. Note the light color of the bones compared to that of the other tissues. The image on the bottom-right shows a 3D visualization of the ankle bones extracted from the scan.

CT technology was originally developed almost 50 years ago and is currently very advanced. The latest machines have an in-plane resolution between 0.20 and 0.25mm, capturing 12 bits of data per voxel (4096 gray levels). Additionally, the intensity values from each individual scanner are normalized into Hounsfield Units (HU), which maps the intensity response of water to 0 and air to -1000. Bone tissue intensity responses are then mapped between 45--3000 HU. Accessibility and cost limitations prevent us from creating new large data sets. Since only hospitals have a need real for them, the scanners are only available in hospitals, where they are reserved for clinical use.

The second, and arguably more important issue, is the safety of the subject. CT scans discharge large amounts of X-ray radiation to the subject. X-rays have the potential to mutate genetic material within the body, which could lead to cancerous growth. Therefore, when doctors order CT scans for their patients, they restrict the region of exposure to the scan. Even if a large number of CT scans were publicly available (you can find a few here), most of them would not image the entire body.

Another common 3D medical imaging modality is Magnetic Resonance Imaging (MRI). Like CT, MRI produces a 3D volume image of the subject. It is based on the response of the body to a strong magnetic field, which affects the hydrogen protons in the water molecules contained in the body. Because it depends on water content, MRI produces more contrast in the soft tissues of the body and is less appropriate for imaging bones. Additionally, their resolution is lower than CT (about 1--2mm in plane) and the scans require more time. However, since they are based on magnetic fields, MRI scans do not expose the subject to harmful radiation. MRI is considered a complimentary technique to CT imaging.

We have spent some time working with CT data to extract bone geometry and will discuss our techniques and issues encountered in detail in future posts.