Home Projects School Valve models from 64-slice cardiac CT scans (using active appearance models)
Valve models from 64-slice cardiac CT scans (using active appearance models) PDF Print E-mail
Written by Administrator   
Tuesday, 29 December 2009 19:46

This project is the basis for my thesis for my masters degree at UConn. My main goal with the project was initially obtaining the data, which was more challenging than anticipated. Hospital PACS systems are generally designed to store lots of imaging data and make it available for 30 days or so. Then the images go into an archive and can be retrieved when necessary... which generally means one patient at a time. PACS is not designed to retrieve large batches of data, like the 900+ patients we got IRB approval to analyze. The 125 or so patients that I did get will be very useful for future imaging studies though.

I was interested in active appearance models (AAMs) as a method to extract 3D valve models from the CT data because of a project I did in my first semester in a pattern recognition class. I actually used active shape models (ASMs), which are very similar to AAMs, to analyze 2D brain images. It worked fairly well and was able to extract the features from most target images. It wasn't terribly robust because it didn't take into account the texture information like AAMs do. I was hoping to extend the 2D ASMs into 3D AAMs for my thesis.

Transforming the 2D statistical shape model (SSMs) algorithm, of which AAMs and ASMs are subset, to 3D is not complicated. A decision must be made to use polar coordinates, quaternions, or another method of rotation in 3D space, but the general equations and algorithm are very similar. The difficult part is in creating the training data. Creating landmarks on 2D training data might involve picking 25, 50, or more points for each training image. The position of the points is generally the same in each image, so loading a 'master' set of points onto each image and moving the points individually to each landmark is relatively easy.

When trying the 2D approach in 3D it gets a lot more complicated. A simple solution, which has been done in the past is to reslice the volume and apply 2D AAMs to each slice. The drawback to this method is that the 3D objects are of different heights and can occupy different numbers of slices. This can be overcome by normalizing the heights, however the internal structures will appear on different slices. So that method is out.

The method I initially chose was to reslice the CT volumes along the long axis of the aortic valve (the valve were interested in), and establish a set of landmark points that describe the important details of the valve. The points would be defined individually for each of the training datasets and manually moved in a 3D viewer to correct locations. The problem with this is the size of the landmark set. Two hundred points might be necessary to accurately describe the valve shape in 3D, however this is an enormous number of points and would extremely time consuming to click them all by hand.

So now I search for better landmarking methods, and better methods are not limited to using AAMs. I may continue with AAMs, but if I find a simpler feature extraction method, I will probably experiment with it.

 
Copyright © 2017 gbook.org. All Rights Reserved.
Joomla! is Free Software released under the GNU/GPL License.