Abstract:

A novel method for the segmentation of 4D MR cardiac images is introduced in this paper. The method improves the traditional active shape model method by adopting an 3D spatially hierarchical expression of the shape model, which is used as an internal regulation force during the segmentation process. Generation of the landmarks for constructing the shape model is based on the active surface method itself utilizing the long range image force, gradient vector flow (GVF). For constructing hierarchical statistical shape models, initial landmarking is done on a manually segmented training set with different spatial resolutions. Principal component analysis is then used to derive the hierarchical expression of the shape model. Experimental results for 4D MR cardiac image segmentation are presented.

Abstract:

The purpose of this research was two-fold: (1) to investigate the properties of statistical shape models constructed from manually segmented cardiac ventricular chambers to confirm the validity of an automatic 4-dimensional (4D) segmentation model that uses gradient vector flow (GVF) images of the original data and (2) to develop software to further automate the steps necessary in active shape model (ASM) training. These goals were achieved by first constructing ASMs from manually segmented ventricular models by allowing the user to cite entire datasets for processing using a GVF-based landmarking procedure and principal component analysis (PCA) to construct the statistical shape model. The statistical shape model of one dataset was used to regulate the segmentation of another dataset according to its GVF, and these results were then analyzed and found to accurately represent the original cardiac data when compared to the manual segmentation results as the golden standard.