![]() | Only 14 pages are availabe for public view |
Abstract CardioVascular Diseases (CVDs) are considered the number one cause of mortality and the major health concern according to recent statistics worldwide. In Europe, nearly half (47%) of all deaths are from the CVDs. About 85% of overall mortality of middle- and low-income countries is due to the CVDs too. Accurate early detection can effectively reduce the mortality rate caused by the CVDs. Left ventricle of the heart is very important in early detection and diagnosis of CVDs. Image processing in the field of biomedical analysis plays an important role in the detection of CVDs through the cardiac segmentation. To determine the shape of the left ventricular cavity, cardiac image segmentation is performed. Many studies and researches introduced several segmentation techniques. Image segmentation techniques can be divided into two main groups. The first group contains the classical techniques for image segmentation and the second group contains the random walk segmentation techniques. This thesis tackles the cardiac segmentation problem. Implementation of different segmentation techniques has been carried out and comparative studies are done to determine the most accurate and fastest segmentation techniques. The comparison is done by evaluation of the quality and speed of segmentation. The quality is measured using similarity Dice Metric (DM) coefficient, Peak Signal to Noise Ratio (PSNR) and HauSdorff distance (HS). The average execution time is measured in each technique to determine the speed of execution. Three Dimensional (3D) multi-slice short axis Cardiac Magnetic Resonance (CMR) database for different case studies is used to demonstrate the implementation results of the different segmentation techniques. A comparative study has been carried out on the first group of different classical segmentation techniques. The classical segmentation techniques can be categorized as: edge based techniques such as Caselles technique and region based techniques such as Bernard-Friboulet technique, Li technique, Lankton technique, Shi-Karl technique and Chan-Vese technique. Edge-based techniques such as Caselles technique gives better results when the initialization step is suitable and the image has high intensity gradient at the edge between the objects in the image. Wideband region-based techniques such as Bernard-Friboulet and Li give worse similarity results as they tend to oversegment the image and take much more time. Narrowband region-based techniques such as Shi-Karl, Chan-Vese and Lankton- Yezzi algorithms are not disturbed by the presence of bright region on a wide range region far from the initial contour or by the smooth variations of intensity inside the object. The image results of Lankton-Yezzi segmentation technique is sensitive to Abstract IV initialization. Shi-Karl, Chan-Vese techniques give good segmentation results in term of both the similarity and computation time. Shi-Karl technique gives the best results among the first group of segmentation techniques in similarity and execution time. In this thesis, another comparative study has been carried out by implementation of the second group of segmentation techniques. Random walk is a multi-label image segmentation technique that is based on graph-theoretic electrical potentials. Random walk technique for image segmentation gives better results in terms of both the similarity and computation speed than Shi-Karl segmentation technique. The second group contains three random walk techniques; which are Basic Random Walk (BRW) with seeds technique, High Speed Random Walk (HSRW) with precomputations technique and Extended Random Walk (ERW) with priors technique. Random walk techniques depend on a small set of marked pixels or seeds. The Experimental results show that BRW technique with seeds gives high quality results in medium speed. ERW technique results are the most accurate segmentation technique because it takes in consideration more characteristics about the regions due to the incorporated priori and the execution of segmentation has medium speed. HSRW technique gives high quality results are close to the corresponding similarity measurements in basic random walk technique with seeds. The pre-computations in HSRW offline mode reduce the execution online time. HSRW technique performs the segmentation process in very high speed. A novel segmentation technique for cardiac image segmentation is proposed by mixing the characteristics of the HSRW with pre-computation model and ERW with prior model to improve the segmentation. The speed of segmentation in the proposed technique is very high. The proposed technique for cardiac segmentation is a robust and very accurate technique for the delineation of the Left Ventricle (LV) endocardium and epicardium to segment the cavity and myocardium of LV of the human heart. The LV performance parameters such as LV volume in diastolic and systolic phases, Stroke Volume (SV) and Ejection Fraction (EF) of the LV can be estimated from the segmented images. This technique considers the intensity values and performs efficiently with blur images and weak edges. The LV performance estimation is based on the accuracy of the segmentation and it helps the doctors in early detection and monitoring of CVDs. |