*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Integrated 3D Anatomical Model for Myocardial Segmentation in Cardiac CT Imagery
Committee:
Dr. Anthony Yezzi, ECE, Chair, Advisor
Dr. Patricio Vela, ECE
Dr. John Oshinski, BME
Dr. Marina Picinelli, Emory
Dr. Brandon Dixon, ME
Abstract: The objective of this thesis is to derive, present, and validate a novel multi-dimensional automatic edge detection algorithm based on shape priors and principal component analysis (PCA) for the challenging and clinically relevant task of myocardial segmentation in coronary computed tomography angiography (CCTA) images. We propose a highly customized parametric model with implicit representations of the segmenting boundaries (3D) of the left ventricle (LV), the right ventricle (RV), and the epicardium (Epi). We develop shape prior models for each anatomy by performing PCA on a set of manual segmentations. We estimate a set of 3D pose and shape parameters (weights corresponding to each principal component) for each of the three anatomical shapes by fitting the models to a customized image appearance model optimized via iterative gradient descent approach. For modeling the image appearance, we generalize the Chan-Vese image segmentation model which treats the image as a piece-wise constant function while enforcing a strict ordering of the image statistics of different regions. We develop a dynamic adaptive background model which treats the complex background region as binary clusters with both very low and very high image intensities. We use all three shape priors simultaneously by coupling them through an overlap and nearness penalty term in our energy functional, thereby constraining the regions to be non-overlapping. The presence of both pose and shape parameters with modeling constraints leads to a challenging optimization problem. We develop a novel automated, shape-adaptive way to choose the gradient parameters weighting dynamically during the fitting process which enables successful convergence. To further customize the models to cardiac anatomy, we modify our pose model to a blended 2D version with a set of separate 2D poses, one each at the anatomical base and apex, and generate a blend of the two poses for all the intermediate slices. This allows us to explicitly capture the torsion and shearing present in the cardiac anatomy leading to better results at the apex and septum. Finally, based on the observation that the RV intensity changes from bright to dark (or vice-versa) from the apex to base, we formulate a piece-wise linear image appearance model specifically for the right ventricle. We test our formulations on a set of 45 short-axis cardiac CCTA images and show clinically acceptable results.
