Toward fully automated analysis of tagged and nontagged MR cardiac images

This paper presents three approaches to the problem of obtaining the left ventricular boundaries from cardiac MR data. The first presents a new model based approach for the detection of the endocardium from 4D MR cardiac images. The method proposed here links shape modeling and edge detection to provide a compact representation of the endocardium. A spatio-temporal edge detector has been designed to incorporate the temporal information available in 4D images. This edge detector has a stronger response to dynamic edges than static edges. Since the ventricle is a dynamic shape, boundaries detected using this edge detector are far better than those detected using a spatial edge detector. The output of our edge detector is iteratively corrected using a spherical harmonic model. This model based approach allows us to overcome the problems of noise and missing boundary information. Our system is fully automated and its output consists of the extracted boundary in each slice and a 3D surface model for each time instant. Quantitative evaluation is done by comparing the results of the algorithm with manually extracted ground truth for 12 data sets. The second approach uses filters applied across the detected tag lines to remove the tags from SPAMM-tagged MR data to allow existing boundary detection algorithms to function with minimal changes. The third approach uses the Fuzzy c-Spherical Shell algorithm directly on tagged (and untagged) data to determine the approximate LV center.