|M.Sc Student||Cohen Nadir|
|Subject||Semi-Automatic Segmentation of the LV Cavity in Contrast|
Enhanced Echocardiographic Clips
|Department||Department of Biomedical Engineering||Supervisor||PROFESSOR EMERITUS Dan Adam|
|Full Thesis text|
Myocardial contrast enhanced ultrasound (CEUS) is gaining popularity as an imaging method for assessment of left ventricle (LV) volume and function, and for myocardial perfusion estimation. Those parameters may help diagnose cardiac diseases, especially when diagnosis is difficult. The contrast agent flows through the cardiovascular system, fills up the LV cavity, and improves the visualization of the endocardium and the assessment of LV volume, ejection fraction and other myocardial parameters. Segmentation of the LV myocardium using CEUS is a primary yet important stage in variety of these LV analyses.
Though the contrast agent is used in order to enhance visualization of the endocardial boundary, it should be noted that the flow of the microbubbles within the cavity creates a gray-scale pattern that is continuously changing. Thus, there exists a paramount need for accurate automatic segmentation, as it may help physicians estimate more objectively the function of the heart while performing an ultrasound exam.
Thus, a segmentation algorithm of the LV endocardial boundary is here presented, while using 2-chamber view CEUS clips. The algorithm is applied only to the diastolic frames of the clip. The user defines a ROI and marks three points on the first frame of the clip - the center of the cavity and the two basal points. Location of the basal points in the following frames is found automatically by cross-correlation. Then, attenuation correction is performed on the images, to correct the intensity attenuation along the LV cavity. An initial estimation of the boundary is found using an algorithm based on minimization of an energy function. The final and major part of the algorithm is the correction of the initial segmented boundary, a process that consists of three stages: (1) Endocardial boundary estimation: The estimation is based on several constrains, such as local intensity maximum, intensity level and proximity to previous points of the boundary. (2) Basal area correction: Correction of the boundary there is according to the estimated basal points. (3) Papillary muscle correction: the area of the papillary muscles often has a poor image quality, which requires an additional correction done by a curve fitting of the boundary.
In order to validate our results, we have compared the automatic boundaries to the manually segmented boundaries, in 276 frames of 9 clips taken from 9 patients. The measured mean absolute distance (MAD) between the two boundaries was 5.9 4 (SD= 1.40 ) pixels, out of a total boundary length of 838.58 (SD= 116.5 5) pixels, which represents an error of 0.71 % (SD= 0.14 %). In addition, we have compared the areas defined between the two boundaries. The mean area of difference (between the 2 boundaries), divided by the cavity area of the boundary drawn manually was 9.90 % (SD= 2.09 %), and the mean Jaccard index was 0.90 (SD= 0.02 ). These results prove high similarity between manual and automatic curves and areas.
Our algorithm provides a semi-automatic and accurate segmentation, which enables an objective and efficient segmentation of the LV cavity. This segmentation algorithm could enable myocardial perfusion measurements, which could improve coronary artery disease diagnosis.