Segmentation of the Visible Human Data Sets

Debra Quackenbush, Peter Ratiu, John Kerr

One of EAI's goals in utilizing the Visible Human data sets the National Library of Medicine's Visible Human Project is to create an accurate three dimensional computer generated model of the normal human anatomy. The first major task in achieving this goal is the segmentation of anatomical structures. While automatic or semi-automatic segmentation is a major preoccupation in the biomedical research community, it poses obvious difficulties. Many researchers are hoping or waiting for the advent of automatic segmentation.

True automatic segmentation can perform the task based on color differences between structures. For instance, we have the capability to save the pixels within the range of the rgb values corresponding to the striated muscles, write them out as cross sections, and automatically fit a curve to the object on each frame. However, in the majority of the cases, this method cannot find boundaries between different muscles. When it comes to other tissue types, like bones, these cannot be consistently distinguished from fat and connective tissue, let alone periosteum.

Other, shape based automatic segmentation algorithms are also in their heroic stages only, tackling tasks such as segmentation of the eyeball or the kidney. Perhaps, the best approach for a future automatic segmentation will be based on the preexisting knowledge of anatomy. In order to "teach" the computer anatomy one has to first painstakingly perform the manual segmentation on all the approximately 50 gigabytes, on the 6300 hundred or so sections of the two Visible Human data sets. We will have to create a knowledge base of the shape, orientation, size, normal variations of each and every structure. It will be the task of computer scientists to translate this into information that the computers can understand and use. But in order for them to succeed, we will have to provide them with accurate empirical data. And although there are some good references for cross sectional anatomy, the work that has to be performed is groundbreaking in its level of detail and tediousness. The limited number of good cross sectional anatomy references also have their flaws. First, the majority of these references are in the transverse plane, with a very limited number of books addressing the sagittal plane and only a few scattered views of the frontal plane. Secondly, these references merely point to different anatomical structures shown in the cross section but do not encircle them to show the respective borders in relationship to the other surrounding structures. Finally, these references also seem to skip over certain sections of cross sectional data which prove to be difficult in determining accurate anatomical identifications.

Another problem that we found ourselves running into, and we are sure many of our colleagues have too, was the overwhelming amount of data one has to deal with. The Male data set is referred to as occupying 15 GB disk space, and the female 40 GB. This can be reduced to circa 3.5 GB and 9 GB respectively, by eliminating the blue background and replacing it with a black one, which occupies virtually no space in run length encoded (RLE) images. Still, a vast data set, yet in a minute our situation will look even worse. Since many of the structures show up better in planes other than axial, we had to reformat the data sets in the other two conventional planes - coronal and sagittal. Now we have up to 11.5 GB for the male and 27 GB for the female. For more detailed studies, such as the ones we are currently performing on the head and neck of the two specimens attempting to research new findings we found that more reformatting is necessary. We have reformatted the head and neck region of the female in non-conventional planes, at 15, 30, 45 and 60 degrees offset from the coronal plane about the z axis. (figure 1.) This cannot be performed in the Male data set, due to the fact that the voxels anisotropic and the very significant interpolation would blur the data. But for the female data set, we have now 63 GB

Our group has successfully completed the segmentation of the organs and systems in the Visible Human Male data set and partially of the female data set through interactive manual segmentation. We have utilized and developed a segmentation tool as part of VisModel, EAI’s commercial modeling software. VisModel is a segmentation tool which fits NURBS curves to the interpolation points laid down by the person performing the segmentation.(figure 2) The curves can then be edited by changing the position of the interpolation points, by deleting or by introducing more control points (figure 3). The interface allows for associating names with the different sets of contours, thus allowing for simultaneous segmentation of several structures (figure 4).

The resulting segmentation data can then be used to create surface geometry, volume rendering, as well as for quantitative analysis of the anatomical entities. Surface geometries can be used as the basis for virtual reality applications in the medical field. Volume rendering has been engaged in creating the Dissectable Human CD-ROM, released in June 1996 by Mosby-Year Book, Mosby’s Systems Atlas of Human Anatomy, published in October 1996 and the Regional Atlas due to be in early 1997. Volume rendering and EAI’s Dissectable Human CD ROM has been specifically addressed in John Kerr’s paper.

Anatomical segmentation and three dimensional anatomical reconstruction can also be used in visualizing anatomical structures that have not been previously described and are still controversial. Many of these structures have been visualized using planes other than the original transverse plane of the data set. Through the use of the computer the original data set can be re-cut into other standard planes such as the sagittal and frontal planes. This aids in the visualization of these new anatomical structures due to their unorthodox configurations and in situ relationships One such structure is the sphenomandibularis muscle, which has been identified in both the male and female data sets in a collaborative effort with researchers from the University of Maryland School of Dentistry. We have, also in collaboration with the University of Maryland School of Dentistry identified and segmented the dural connection of the rectus capitis posterior minor muscle.

In the course of the segmentation process we have identified a number of peculiarities in both the male and female data sets. Some of these are normal anatomical variations, that can be referenced in widely accepted human anatomy textbooks and atlases such as Netter’s Anatomical Atlas and Gray’s Anatomy. In the male data set, different venous patterns are represented in the right and left forearm. The right forearm presents a median cubital vein (figure 5. ) whereas the left forearm presents an intermediate basilic and an intermediate cephalic vein located in the same area in place of the median cubital vein (figure 6. and figure 7.).

The heart angle and position in the male thorax is horizontalized relative to normal cardiac position.

Other variations that we observed are not documented in the literature. In the male data set, the internal and external jugular veins do not separately branch off of the brachiocephalic vein as described in the anatomical texts. Rather, these veins arise from a single, common vessel that we have identified a common jugular vein (figure 8.).

A third group of findings are due to the pathological changes in the specimens as well as post mortem artifacts. In the male, there is a significant amount of atheromatous plaque buildup in the aorta and common carotid arteries (figure 9.). At levels, the plaque buildup occludes almost three quarters of the lumen. There are several sections in the male data set, such as the left hand, that contain hematomas. In these areas there is a loss of venous structure and visual obstruction due to the clotted blood (figure 6.), which makes segmentation of the superficial veins of the hand difficult.

The male also presents a doliocolon, an abnormally long colon which therefore forms an extra loop at the hepatic flexure (figure 10.).

Only the left testicle is present (figure 11.) and no evidence of the remains of a undescended testicle within the abdominal cavity. However the intraabdominal portion of the spermatic chord can be identified, which suggests that the testicle was removed after its descent.

The cerebellums of both specimens, but especially that of the male presents a massive herniation through the foramen magnum into the spinal canal and through the intervertebral space into the space between the recti muscles (figures 12, 13, and 14)

Both, the male and the female specimens were frozen with their feet in plantar flexion and inversion. Their arms and hands frozen in pronation and flexion so that accurate segmentation of both the hands and the feet in any one cross sectional plane is extremely difficult (figure 15.).

In the female data there is a marked hypertrophy of the left ventricle. There are numerous pericardial/myocardial adhesions along with scars consistent with old myocardial infarctions. There is severe pericardial thickening which has significantly interfered with the segmentation of the pericardial sac (figure 16.) The female also exhibits atheromatous plaque buildup in her aorta but to a much lesser extent than the male.

There is a space occupying mass displacing the spinal cord in the thoracic region (figure 17.). This could be seen as a problem in future segmentation of the spinal cord and its surrounding structures due to the abnormal position.

There is a 3cm x 2.5cm x 5cm tumor in the region of the rectum. The tumor attaches to the posterior wall of the rectum through a very small base, which suggests it is a pedunculated polyp (figure 18.).

Although the segmentation of the Visible Human has been successful, a few problems exist involving the segmentation of the specific anatomical entities. Segmentation of the vascular system was extremely difficult due to its involved branching (figure 19.). The following of the secondary vessels off main identifiable vessels helped in properly identifying the smaller branches. Yet many of the smaller branches could not be accurately identified nor properly segmented and connected to the main branches. All the vessels present in one cryosection plane could not be segmented all at once due to this hierarchial approach of segmentation (figure 20.). Dedicated software such as VisModel can considerably speed up the tedium of segmentation but the time factor still remains an important issue.

The question of boundaries also arose in several different regions. Such is determining the separation between the myocardium of the left ventricle and that of the right ventricle. This has been performed through educated decisions by the anatomist since no objective boundary can be identified on a macroscopic level. These educated decisions were also carried over in deciding the boundaries between the processes of the vertebrae.

The segmentation of the spinal muscles also proposed a similar problem due to the blending of the individual muscles together. Due to the blending of the muscle fibers, accurately separating them into individual identifiable muscles was an unaccomplishable task. Virtual spaces such as the inguinal canal have proved to be extremely challenging entities to segment.

Segmentation of the Visible Human data sets offers many additions to the original goal of a three dimensional computer generated anatomical model of the human body. With this extensive segmentation of anatomy, questions will also constantly arise and answers constantly be discovered on how to realistically and accurately represent the Visible Humans.