Saturday, October 21, 2006

More segmented left-atriums

I have been experimenting to find out how my implementation performed in segmenting the atrium. Here is one of the cases I have been looking at very closely. I especially like this case since the original image from which the left atrium needs to be extracted is quite 'mesh'-ed within the MRI with the atrium concealed completely inside. Here's the original image:


This is almost the entire thoraic region segmented from an image which was obtained by subtracting a post-angiographed from a pre-angiographed MRI (in order to remove the bone structures).

The atrium, after computing the local maximas and basic components surprisingly yielded very few orphans (those which dont belong to a basic component). This is however, not the case with some other segmented images/MRI. I quite do not understand yet what feature of an Euclidean distance transformed image causes these orphans to occur. However, a very preliminary speculation has led to me thinking that it might be because of the absence of local-maximum points in the vicinity of image-edges. Usually, it is the case that when searching for a voxel's basic component close to an edge of the image, the path gets led to a neighborhood of voxels with 0 EDT values. Once it enters such a neighborhood, the search terminates, since the path has exited the EDT-transformed image and it's impossible to search any further. There could be ways of somehow searching further by querying perhaps the original image itself. I might be soon looking at these possibilites soon.

Here is a segmented atrium I obtained by running the implementation on the original image above. The good thing about this segmented atrium is that it reveals all the pulmonary-vein drainages, in addition to also showing how the drainages are branching out.



No comments: