UC Irvine

Echocardiography has been the preferred imaging modality to study the heart chambers for routine screening purposes. However, it suffers from subjective interpretation and inter- and intra-operator variability. The clinical need for fast, accurate and automated analysis of echocardiograms is currently unmet.

Precise automatic analysis will offer considerable improvements in clinical workflow efficiency and reproducibility. Automatic segmentation and tracking of cardiac chambers are the essential steps for consistent calculation of clinical indices such as ventricular volumes and ejection fraction. Currently, segmentation of echocardiographic images is a manual and tedious task. Machine learning has recently been the center of attention in radiology helping automatic identification of complex patterns to make clinical intelligent decisions. My contributions in this thesis to the field of medical imaging will embrace: (1) Fully automated approach for segmentation of heart chambers in echocardiography images, based on discriminative deep-learning algorithms. (2) Improving segmentation performance by taking advantage of adversarial training. (3) Fully automated tracking of all heart chambers in cardiac cycles using semi-supervised deep learning methods.

The functionality of our approaches is evaluated using a dataset of 1000 annotated images from 100 normal subjects’ echocardiography records. Using several measures of errors, the degree of similarity between the manual and automatic segmentation was defined. Our promising results corroborate that deep-learning algorithms can be successfully employed to solve the challenging problem of automatic echocardiography analysis.