Abstract
The echocardiography examination of the heart provides a proficient assessment of the cardiac functional performance. The scanning of the heart is either recorded in an archive of analog videotapes or output in digital DICOM format that requires intensive manual annotations for the delineation of context-specific information within the echo. In either case, it is challenging for clinicians to search and browse the contents for diagnosis purposes. The focus of this dissertation is to present an original and fully automated approach for the analysis and quantification of echocardiogram video archives. The implemented system has as input the echocardiogram studies digitized from the input analog videotapes or digital echoes in DICOM un-annotated format, and eventually provides fully identified echo views with the corresponding cardiac chambers labeled. This research provides the following main contributions: An adaptive partitioning technique was developed for defining the temporal structure of an echocardiogram study through the detection of the boundaries separating different views/mode at any transducer change in angle and/or position. The developed technique is based on selective weighting and multi-step thresholding of edge and intensity based measures that are tailored to the properties of the echo studies. Experiments performed on analog videos produced satisfactory precision and recall results and a boundary detection accuracy of 99%. A fully automated multiple-cavity segmentation technique was then developed based on discriminant analysis and morphological watershedding for the delineation of the cardiac cavities. Linear discriminant analysis is utilized for binarization via ternary thresholding of the echo intensity levels as a base for filtration and noise reduction. A method is then presented that provides a special marking technique to overcome the oversegmentation downside of watershed segmentation. The correlation coefficient of the ground truth values and the segmentation results is shown to be 0.88 and is then further enhanced after object labeling and splitting of under-segmented results.A technique is introduced to label the segmented objects and overcome the occasional under-segmentation in the segmentation module due to the challenging nature of the ultrasound intensity levels. The proposed approach adds on the segmentation module by incorporating region splitting and deformable-model guided labeling of cardiac cavities. A cavity labeling framework is developed to model the properties of the main cardiac chambers through learning the properties of parametric deformable templates fitted to various labeled cardiac chamber samples. The trained statistical models are then utilized to evaluate region splitting hypotheses for the undersegmented results obtained from the segmentation module. During this stage, deformable model fitting, region splitting and consistency checking are executed simultaneously. The likelihood of region splitting is evaluated using a cost measure based on the strain deformation energy acquired after fitting the trained deformable models to the tested data. Once region splitting is performed, the learned statistical model is directly applied for cardiac cavity recognition and labeling. This work also provides a multiplicity of quantitative cardiac measurements, including left ventricular volume, ejection fraction, cardiac output and stroke volume. Performance of object labeling is evaluated within analog and digital contexts and proves superior robustness against extension of the echo library and within an all-digital context to accommodate for current and future echo libraries. Object splitting adds a significant enhancement to the segmentation results where the correlation coefficient between ground truth and segmented data increases to 0.92. Finally, echo views are recognized by applying a graph-based representation within a fuzzy framework to accommodate the uncertainties inherent to the echocardiogram views. Performance is cross-validated using multiple view models and proves to deliver high recognition accuracy for the tested views. To the best of our knowledge, this is the first dissertation addressing the analysis of the echocardiogram videos for the purpose of indexing their content. Our work is validated on echocardiogram studies of real patients.