Current state of the tongue model

This site attempts to give a visually guided overview of a former project at the Research Laboratories of Electronics, which is part of Massachusetts Institute of Technology. This study, conducted by Reiner Wilhelms-Tricarico, is centered on implementation of a biomechanical model of the human tongue.

The plan behind is to provide a computational simulation of biomechanical and as far as possible physiological features of the structures of the vocal tract. To give an idea of this, view the following two animations generated by an older model that I consider now obsolete ( see: Wilhelms-Tricarico, R., Physiological modeling of speech production: Methods for modeling soft-tissue articulators J. Acoust. Soc. Am. 97 (5), 1995, pp. 3085-3098). These simulations were done with a rather primitive 42 finite element simulation. In the first movie, the transveralis and verticalis muscle in the tongue and the genioglossus and styloglossus are contracted simultaneously, and in the second movie only the hyoglossus is contracting. In both cases, mirrors were simulated so that the back or underside of the "tongue" can as well be seen.

(depending on the browser this short mpeg movie is shown once or repeatedly, for windows it may require the installation of an mpeg player software)

Updated June 2005: This project is no longer continued. I attempted to publish some of the results in the J. Speech and Hearing and Language Research but the manuscript was rejected. Here is the old submitted manuscript (about 2.4 MB): JSHLRmanuscript.pdf. End of update.

Above shown are two views of the current appearence of the model. The large building blocks are hexahedral or prismatic and can subsequently be subdivided as is shown further below. They are delineated by splines, and their shape is hand-edited using a rather awkward program that I wrote mostly in Matlab with some C subroutines. At first glance these objects look like the famous Belgian chocolats that one can buy all over the US in shopping malls at Easter, but then it might just look like chicken meat hunks - way less poetic. Finally the gruesome reality: It's human! The surfaces of the model are rendered with the pixel values that were obtained from the Visible Human project, and the shape of the model follows roughly the morphology shown in the data. So we see here essentially human anatomy obtained from the frozen cadaver of a woman who gave her body to science. The mandible bone (jaw) and some teeth are visible in this rendering.

For a clearer idea of the shape of this model, I include subsequently six different views from orthogonal directions: (on the first row, from left to right: view from back, below, front. On the second row, from left to right: view from left, view from right, view from above.) For larger views, one can click on each image.

After subdividing the above building blocks, the lowest resolution model that can be generated is shown below in several views. The colorful version shows a, merely practical, grouping of the original building blocks and how they are represented by the finite elements.

The finite element model presented has 250 elements at this point. Each element is of the tri-quadratic type and has 27 local nodes: Following the topology of a cube, there are 8 corner nodes, 12 nodes on the edges, 6 nodes on the surfaces and one node in the center (the so called bubble node). In the computational implementation the bubble node will be made superfluous and it is planned to use a 26 node element.

It may be noticed that the representation of the tongue and floor of the mouth does not accurately follow the boundaries of muscles and bones. In most cases this was done to allow a more undistorted shape of the finite elements. But also, to "correct" some of the non-typicality of the original data. For example, the specimen has many teeth missing, which resulted in a unusually expanded tongue blade. So I filled in the teeth gap a little bit. Further, I do not care so much about individual teeth in this model, which has anyways too low a spatial resolution to be relevant for the collision of the tongue with one tooth (however, it may have some relevance for the collision of the tongue with a whole row of teeth). Of course, the inaccuracy does not end here. For example, the whole region of and around the hyoid bone is represented by a few finite element blocks that will have higher than usual stiffness to account for the low deformability of that region. In future simulation studies I'll see how good or bad this works.

One may wonder why the boundaries of the elements do not, in most cases, follow the structure of the tongue. It would be desirable, for better numerics, to model each muscle in the tongue seperately and the layers of other types of tissue between the muscles by their own finite element representation. I had tried this to some extend before and failed to achieve a feasible model that whould not have been represented by many thousands of elements (below I show an example of an incomplete model in which I tried to model muscles individuallly, more or less). Since I did not have any appropriate mesh generation software to aid this goal, I reduced the complexity drastically by allowing the finite element mesh cross any type of boundary between different tissue types. As result, the mesh is built to follow the external shape of the tongue as well as the skelettal boundaries whereas individual muscles are first represented by large blocks that are independent of the finite element mesh and then fiber directions are extracted from these blocks. Here are some of the muscle block representation in a rendering that may lure some bikers to this site:

Not a motor cycle, a collection of blocks from which the muscle fiber directions at each point can be obtained by computing a tangent to the curvilinear coordinate systems in each block.

Link directly to the various muscles in the tongue/floor of the mouth (if underlined, the links should actually exist):

Genioglossus (anterior)

Genioglossus Posterior

Styloglossus

Hyoglossus

Geniohyoid

Digastric

Transversus

Verticalis

Superior Longitudinalis

Inferior Longitudinalis

Mylohyoid

Above: Combination of a midsagittal cut and an oblique cross section that runs roughly parallel to the insertion points of the Mylhyoid muscles to the mandible. Length sections of the Genioglossus muscle are visible (dark matter in the center of the jaw), and cross sections of the Mylohyoid (directly next to the bone of the mandible) and the Hyoglossus muscle (very dark, interrupted semicircle around the tongue body). The tongue tip and the lips are also visible in this iimage.

Below: The second above picture shows an earlier attempt to build a model that directly follows the structures inside the tongue. Different tissue components are color coded. This attempt was abandonded because it always became too complex and a lot of closure problems occured that could only be resolved by inserting strongly distorted elements.

In this model, it is necessary that the directions of muscle fibers are defined in arbitrary orientations relative to the local finite element coordinate systems. The muscle fiber directions can be obtained quantitatively by looking at various cross sections (see above pictures) of the Visible Human Date set and by studying the anatomy in good books. Some kind of middle way had to be found that avoided the total complexity of the real data set.

The direction of muscle fibers is not in all cases visible in the data set. In some cases it can only be concluded from a knowledge of the exterior shape of the muscles. For example, the striation of the genioglossus (under the tongue body, pulling the tongue forward and down) is quite obviously visible. On the other hand, the fiber directions of the geniohyoid which connects a lower point of the mandible (behind the chin) to the front of the hyoid bone is not visible and can only be assumed to be roughly parallel to its outer boundaries.

The specification of muscle fibers is facilitated if a model can be found that descibes the fiber directions (and densities) throughout the muscle and avoids the specification of many individual fibers. For this, I constructed block representations of muscles that are independent of the finite element mesh and capture the shape of the muscle so that the fiber directions are directly computed based on the geometric model. An example is given for the Hyoiglossus muscle. This muscle extends from the sides of the hyoid bone upwards towards the tongue body. It appears to wrap partially around the tongue body. Its function is to pull the tongue down and back. I represented the muscle by four hexahedral blocks and four prismatic blocks. The parametrization of each block allows the computation of tangential vectors (curvilinear coordinate lines) which is then used to obtain (of better represent) the muscle fiber directions.

Above rendered in red is shown the method of representing muscle fibers. The Hyoglossus was represented by prisms and hexahedra and fiber directions were generated by computing the tangential directions of the curvi-linear coordinate lines of the solids. Remark: The orientation (of the blue versus the magenta lines) is irrelevant, it just happed to be that I chose two different orientations for the two types of elements. To avoid a possible misunderstanding - The little blue and magenta lines are not supposed to represent muscle fibers as such. They only visualize the fiber direction at a grid of points throughout the muscle, and they can be computed at any location of the muscle. In addition, the figure rendered in flesh colors is an improved representation of hyoglossus and was actually used to compute fiber directions in the model.

The finite element model constitutes a mesh of points at which the dynamic equations of the model are evaluated in order to compute the movements of the model (which is not yet implemented, just in case you are getting impatient here:-). In order to compute forces that act onto the nodes of the finite element representation, the stress generated by the muscles and all passive stress fields (due to deformation) have to be computed at many points throughout the finite element model domain, and for that the fiber directions have to be known at those points.

To obtain these models, I overlap the finite element model with each geometric muscle model and compute the fiber directions at a large number of points in each muscle. At the points of the finite element model where the fiber directions need to be known, the fiber direction of a point in the muscle geometric model is taken that is the closest to the point in the finite element mesh. Within the finite element mesh, the points at which stresses are evaluated and thus fiber directions need to be known, are the Gauss points used for numerical integration. I use a numerical sheme that is based on 27 nodes and as well on 27 Gauss points inside the elements.

Above: Top and side view for the fiber directions at Gauss points of the Styloglossus. Some of the elements where not drawn to allow the visualization of the fibers. Below, another view of the same representation:

It may be noted that the styloglossus appears here in a slightly exaggerated way. It is generally agreed that there are two main directions of this muscle: It comes from the stylohyoid process at the base of the scull and inserts into the tongue body laterally next to the hyoglossus but roughly perpendicular to the hyoglossus. Some of the fibers interdigitate with those of the hyoglossus and seem to enter the back of the tongue where they can not easily be distinguished from the fibers of the transversalis muscle. Some of the fibers of the styloglossus run anteriorly towards the tip of the tongue and appear to wrap around the tongue body like a sling. When studying the anatomy in the data set, it is difficult to see the styloglossus in many parts of the specimen. Therefore, the actual continuation of the fibers all the way towards the midsagittal plane is a little bit of an exagerration. What is not shown here very well is that these fiber fields are also associated with a density. This density decreases as the fibers approach the midsagittal plane.

The anterior part of the Genioglossus muscle, which pulls the tongue down and forward, is shown in the above two images from different viewpoints. Some of the element were only rendered as wire frames or completely avoided (in the fleeing view) to get a better spatial understanding of the images. I should point out that the "pixidust" that can be seen (if your screen has sufficient resolution) are the many Gauss points within the finite elements, at which the fiber directions are defined.

The posterior part (according to the terms used in speech physiology) of the genioglossus is the not necessarily an independent muscle. (It may be possible to subdivide the genioglossus in even more subunits then just anterior and posterior, but as long as we don't know the accurate "wiring" of the muscles, this discussion doesn't lead to anything).

Above, the Hyoglossus muscle is shown from the left and from the front in the fully wireframed model. The lateral view shows a little potential problem of this kind of muscle representation. On its lower end the hyoglossus muscle ends at the hyoid bone. But where does it end at the other end? Currently, I have assumed that it simply ends as shown but that is only because I can not detect any fibers of hyoglossus in the data. There is a good possibility that this representation is insufficient and that it will be necessary to continue the fiber field for the hyoglossus all the way to almost the tip of the tongue, in which case its course needs to be more or less invented, given the data that I have at hand.

The next muscle presented is the Geniohyoid muscle. It extends from the inner lower mandible to the central parts of the hyoid bone. However, as can be found in the actual data, some of the fibers of the geniohyoid muscle seem to fail to end at the hyoid bone. They pass by the outside of the hyoglossus. So I represented the muscle of two main blocks as seen below. The other picture shows the resulting fiber directions in the FE model.

Above: The red and blue block are the two parts that represent the geniohyoid. A midsagittal cross section throught the Visible Human data set is also shown and two oblique cross sections, one to the left roughtly where the attachments of the geniohyoid to the jaw can be found, and the other about at the front plane of the hyoid bone. Fiber directions are calculated on the surfaces of these blocks and throughout their volume. It may be noticed that towards the end at the mandible (left) the length of these fibers increases. The fiber density is coded by the length of the yellow lines. The other picture shows the resulting fiber directions in the finite element model.

The Mylohyoid muscle is on the underside of the tongue and forms a slab that connects to a rim on the inside of the mandible. The following figures show the modeling and fiber direction modeling of the muscle:

The green slab represents the mylohoyoid. For orientation, three cross sections are shown. To the right is one large and one small cross section through the mandible and to the left a midsagittal cross section can be seen. The images below show the mylohyoid fiber directions in the FE model at the Gauss points:

Only the anterior belly of the digastric muscle is incorporated into the model. It is also questionable if it should be part of the model at all. There was a discussion if it should better be included in a seperate mandible-hyoid model. The way it is implemented in this model is a bit of cheating, since some anatomy books claim that the anterior belly of digastric is connected not in a fixed way to the hyoid bone but in by means of a sliding sling.