Surface Warp Protocol Data Assumptions
Overview
The data is expected to be organized as a collection of subjects. Each subject must have a surface object and a collection of contour objects, which define the anatomy on the surface. The surfaces must have an octahedral triangulation. The contours must lie on the surface. In other words, given a contour and its surface, for every vertex of the contour, there must exist a face (polygon) whose edges or planar region interior to the polygon contain the vertex. This constraint should be enforced by the drawing program that generates the contours. The subjects are allowed to have sets of contours that do not represent the same anatomy. For example, it is allowed for a piece of anatomy to exist in all but a few subjects. For a contour to be included in the atlas, it must be present in more than one subject. A given subject's surface will only be warped to contours that are in the atlas. If the subject does not have some of the contours in the atlas, then those are necessarily excluded from the warp.
Currently, the contours may be encoded to the MINC OBJ contour and LONI UCF formats. The surfaces must be encoded as a MINC OBJ surface.
Directory structure and filenames
There are many ways to structure and name the data used in this process. To avoid case-by-case implementations of the process, the user must rename and organize the data to the form that is expected by the process. Each subject is identified by an arbitrary unique string, and each anatomical contour is defined by an arbitrary unique string. Each subject must have its own directory. In each subject's directory, there must be a surface file, "surface.obj" and contour files, "contour_${id}.obj", where ${id} is the string that identifies a particular piece of anatomy.
For example, in our directory "study", we have subjects named 1000, 1001 and 1002 and contours named 1, A, B, and C:
study/1000/surface.obj
study/1000/contour_1.obj
study/1000/contour_A.obj
study/1000/contour_B.obj
study/1000/contour_C.obj
study/1001/surface.obj
study/1001/contour_1.obj
study/1001/contour_A.obj
study/1001/contour_B.obj
study/1001/contour_C.obj
study/1002/surface.obj
study/1002/contour_1.obj
study/1002/contour_A.obj
study/1002/contour_B.obj
study/1002/contour_C.obj
Warp Model Constraints
This section describes the expectations of the algorithm used to compute the warp. There are constraints on the triangulation and size of the surfaces and the geometric properties of the contours.
The surfaces must have an octahedral triangulation with 65538 vertices. This surface can be imagined the following way: consider an octahedron; each face of the octahedron is a triangle; for every face, add vertices that bisect the edges of the face and add edges connecting these new vertices, creating four faces from every face; apply the previous step N-1 times recursively. The number of steps N is the order of the triangulation. The number of vertices is 22*N + 2. The number of faces is 22*N + 1. The implementation expects N = 8. The implementation also expects a specific vertex and face ordering; however, this is beyond the scope of this documentation.
The contours can have properties that prevent a well-behaved solution. In general, the warping algorithm computes a transformation that is an elastic homeomorphism. There are some situations where anatomical landmarks will constrain the warp such that the solution is not a homeomorphism. Most of these situations can be avoided by proper definitions of the landmarks. Here is an enumeration of the conditions that can cause such an issue:
- duplicate landmarks
- two distinct reference landmarks are mapped to the same target landmark and vice-versa
- topological inconsistencies
- two points are close neighbors in the reference space and not the target space and vice-verse
- large deformations
- a reference/target landmark pair specifies a deformation that is too large given the behavior of the neighboring landmarks
- arises from the elastic model used to compute the warp
The above situations should be considered when drawing contours and defining the anatomy used for the warp. In the case of duplicate landmarks, two contours must intersect for this to occur. For topological inconsistencies, this occurs when contours form a T-junction, that is, an endpoint of contour A is a neighbor of contour B in a neighborhood that does not include the endpoints of contour B. This is a troublesome condition to avoid. If the T-junction occurs across all subjects, then splitting contour B at the closest point to the junction would solve the problem. However, when the T-junction occurs in only some subjects, the point to split is not well defined. Regardless, anatomical aspects of the intersection should also factor into the handling of this situation. For large deformations, there is little that can be done to the data to avoid this, as it arises from the limitations of the elastic model. If this situation is encountered it would be worthwhile to test a range of elastic parameters.