PDF version of this entire document
Stepping aside from literature, an important problem to overcome will be use of tags or analysis in their absence. One potential way to approach this problem is to assume we get a maximally-spread diamond/square when its furthest two edges are far apart, usually before it gets compressed by movement or localised stress. To demonstrate this point imagine a complete blurring of edges and think of what would happen if you measured the radius of a circle passing through/within (or conversely containing) the diamond. Depending on what corners it passes though, its size can tell us something about the level of localised compression and we have code that calculates this. If put on top of the image a radiologist may be able to assess the level of pressure applied by nearby motion that flattens a diamond. If circles are too messy as annotation, then colour too can be used. We want to explore this idea of visualising a surrogate of stress because we have not seen it in other papers yet.
If a distance or other general measure is used for calculating the
stress at one point in the tagged image, a method is still needed
for identifying the boundaries of the blocks which tagging yields.
Cross-frame analysis is required here and one method for finding a
match amid movement is a shuffle transform based on MSD/SSD. Initial
tagging is just a grid, which is simple to set up. Having initialised
a grid layout (be it diagonal or upright wrt xy), which is
essentially just a board of blocks, we can then apply shuffle map
wherein two frames (or more) are taken at a time and then for each
pixel/voxel (or junction in the tagging, for speed and avoidance of
interpolation) searches for a best match after the movement, using
available information in all 3 axes (shuffling within a cube of fixed
size, where vicinity/neighbourhood is defined and limited by a window
size). From this we obtain simple tagging which is quite resistant
to blurring or can at least estimate best match when the tagging is
blurred. The size of the shuffle window ensures that the grid does
not move too far, too fast (although it can be adjusted to explore
a wider window in particular areas of the image) and to avoid folding
or tearing there can be constraints that need not rely on any splines
which are diffeomorphic, such as clamped-plate splines. The merit
of this approach is that it uses information in three dimensions without
relying on heavy computation (shuffling can be encoded as a 
=64
bit mask), it explores frame correlation, and it is resistant/sensitive
to blurring/noise because of a `best fit' approach. We have begun
implementing this for testing on low-resolution data that we got.
We are exploring a way of utilising tagged cardiac data as means of
improving segmentation and perhaps - if this goes further - statistical
analysis of tags/segmented parts can take place. The nature of the
imaged subject/s does not matter to us; it is just that many hours
were spent in vain trying to find tagged images and the most/best
we found was a single low-resolution (under 300 pixels) sequence of
10 images. Now, back to the literature.
Roy Schestowitz 2010-12-25
