Contour Tracking in Echocardiography via Sparse Modeling

Cardiovascular disease is the leading cause of death worldwide. To reduce the mortality, it is desirable to develop automated quantitative cardiac function evaluation techniques using non-invasive imaging tools. Segmentation and tracking of cardiac borders from echocardiographic sequences play an important role in cardiac functional analysis that estimates important functional parameters, such as ejection fraction, myocardial wall thickening, and myocardial strain. Robust and accurate automatic segmentation of the left ventricle, especially the epicardial border, is very challenging in echocardiography, due to low image quality. The inherent spatiotemporal coherence of echocardiographic data provides useful constraints that can be exploited to guide cardiac border estimation.;This dissertation focuses on exploiting the inherent spatiotemporal coherence of individual echocardiographic data to constrain cardiac contour estimation, instead of learning offline spatiotemporal priors from databases. We propose a dynamical appearance model (DAM) based on sparse representation and dictionary learning for tracking both endocardial and epicardial contours of the left ventricle in echocardio-graphic sequences. The contour tracker is initialized with a manual tracing of the first frame. It employs multiscale sparse representation of local image appearance and learns online multiscale appearance dictionaries in a boosting framework as the image sequence is segmented frame-by-frame sequentially. The weights of multiscale appearance dictionaries are optimized automatically. Our region-based level set segmentation integrates a spectrum of complementary multilevel information including intensity, multiscale local appearance, and dynamical shape prediction. The approach is validated using 2D+t and 3D+t canine echocardiographic images acquired from both healthy and post-infarct canines. The segmentation results agree well with expert manual tracings. Advantages of our approach are demonstrated by comparisons with a conventional pure intensity model, a registration-based contour tracker, and a state-of-the-art database-dependent of line dynamical shape model. In addition, we apply our method to two important cardiac functional analysis tasks: ejection fraction estimation and myocardial strain estimation. We achieve good ejection fraction and myocardial strain estimation results using our segmentation outputs. We also demonstrate the feasibility of applying our method to clinical data.;We also propose an extension of the DAM using stochastic online dictionary learning to alleviate the drift in tracking. It utilizes a stochastic optimization technique and processes a mini-batch of training examples at a time, which results in lower memory consumption and lower computational cost. In contrast to the our base method where dictionaries are trained only using the last segmented frame, the new stochastic online learning procedure optimizes the dictionaries and the corresponding weights by aggregating the information of all the past frames while adapting the dictionaries to the latest segmented frame. We weight the past information to control the rate at which the past information is updated by the new information. This updating rate varies with appearance scale and maintains a balance between the past and the new information. Experimental results show that this stochastic extension effectively improves the accuracy and robustness of endocardial segmentation and computational efficiency compared to the original batch learning procedure.