Robust segmentation of echocardiographic sequences through spatiotemporal planning and by knowledge-based local and global methods

A new multi-stage approach is proposed for automatic endocardial border tracking on sequences of digital echocardiographic images. To cope with noise and variability as well as to reduce computation, an approximate contour is assessed for the initial frame in 4:1 decimated images. A composite image (GZD) is obtained by a pixel-wise AND operator between the zero-crossing and gradient direction image, where gradient is above a threshold. A motion image, obtained by integrating the GZD images over the sequence, is used to limit ROI, where the only edge elements belonging to the LV endocardial border are collected and connected by a local radial search. Then edge segments are 1:4 zoomed, interpolated and refined by an optimal global search, based on a likelihood function, maximized by Dynamic Programming. This step is repeated for the following frames using the previous contour as a guide. This approach was tested on a set of 15 DICOM files obtained by a digital (Sequoia, Acuson) echocardiographic system. Automatically derived contours show differences with respect to manually defined contours within the range of observers' variability.