| Heart diseases are the leading cause of death in the US. Left ventriculogram (LVG) is widely used for diagnosis and prognosis of left ventricle (IV) diseases. End diastole (ED) and end systole (ES) boundaries from LVG are interested in because of their importance of assessing heart diseases. Currently manual tracing is the only reliable way to obtain these boundaries, which is tedious. Automated LVG boundary delineation has been studied for decades, but none have ever been able to compare to manual tracing in terms of accuracy.; The goal of this project is to design some algorithms to delineate the ED and ES boundaries from LVGs so that the ED and ES volumes and ejection fraction (EF) derived fall into the interobserver variability range. In this dissertation, a systematic 3-stage approach is discussed to fulfill the task, including a pixel Bayesian classifier, a shape calibrator and a rejection classifier.; The pixel Bayesian classifier places the pixels in the image into 3 classes, the background class, the ED-not-ES class and ES class. Then ED and ES regions are formed and the initial boundaries are traced along the regions. Those boundaries are then calibrated by a shape regression. The rejection classifier determines whether or not to accept the results. The system processes the ES result conditioned on the ED result and needs the user to input the two aortic valve endpoints and the apex to guide the system.; The class conditional probability and the prior probability used by the pixel Bayesian classifier are estimated nonparametrically. The class conditional probability is estimated with a generalized look-up table (LUT). The prior probability is estimated by aligning the aortic valve angle and long axis of the LV. Moreover, the shape regression coefficients are generalized with cross validation. The rejection classifier is also trained with cross validation. Finally, the whole system is optimized by minimizing the cost function defined in the dissertation.; The system is trained on a database of 375 study cases and tested on an independent 18 cases. The current results shows that the system performance is close to human interobserver variability. |
