Thesis (M.Sc.) - Cairo University - Faculty of Engineering - Department of Systems and Biomedical Engineering

Evaluating the heart global function from magnetic resonance images is based on estimating a number of functional parameters such as the left ventricular (LV) and right ventricular (RV) volumes, LV and RV masses, ejection fraction, and stroke volume. Estimating these parameters requires: First, accurate segmenta- tion of the main cardiac champers, LV and RV. LV segmentation can be done by many techniques with a plausible performance; However, RV segmentation is still a challenging problem due to the complex shape of the RV, presence trabecular muscles and low tissue-to-blood contrast. Second, accurate calculation of the volumes enclosed by the inner and outer surfaces of the LV and RV chambers. Currently, the volume calcu- lation is achieved through acquisition and segmentation of a large number of short-axis (SAX) cross- sections of the LV and RV, which is time-consuming, expensive and causing patient inconvenience. Unfor- tunately, reducing the number of the acquired cross-sections results in undersampling the LV surfaces and hence increases the errors of calculating the volume. In this thesis, we propose two packages of novel meth- ods in both RV segmentation and volume calculation from small number of MR slices. In RV segmentation, an Active Shape/Appearance Model (ASM) based system is introduced for accurately segmenting the RV from MR short axial (SAX) images. The RV shape variations are independently modeled using two separate (dual) ASM models in the Bookstein domain. In order to better segment the low contrast regions at the apex, the method is extended for multiple-2D images where the variations of all cardiac levels are incorporated in a single model. For the volume calculation, new method based on fusing the long axial (LAX) and SAX views of the heart is presented. In this method, the LAX contour is used to swipe the SAX contours to fill in the missed LV surface between the SAX slices. Furthermore, novel geometrical models that combine infor- mation from SAX and (LAX) views are invented. The results from 25 datasets and 5 CT-based phantom showed that our methods outperform most of the used methods in the literature

Issued also as CD