For three-dimensional echocardiography (3DE) to have greater clinical use, there will need to be automated means for estimating cardiac function parameters such as left ventricular (LV) volume directly from the 3DE data set. A prerequisite for estimation of LV volume is the accurate extraction of the endocardium over a cardiac cycle. In this paper we present a semi-automated algorithm that, with minimal operator input, effectively tracks the LV boundary through the spatial and temporal sequences of 2D frames generated by 3DE. Variations in imaging conditions and heart motion make it difficult to develop effective prior geometric and dynamic models for the LV. However, operators can easily locate a few landmark points on the boundary in a given 2D frame. Our algorithm begins with the operator marking some highly visible points along the boundary in a few spatially separated frames at end-systole. This takes a few seconds to complete, and is the only operator input. Full boundary estimates in these initial frames are completed by spline fitting to the selected points. These estimates are used to establish search regions for the intermediate frames at end-systole, within which boundary points are specified as those having highest edge probability. The use of search regions avoids matches to non-endocardial edges. A similar procedure is then used for the temporal sequence of frames at each spatial location: the boundary is tracked by finding points of high edge probability within search regions initialized by the end-systole estimate at that location. LV volume as a function of time is then calculated from the set of estimated boundaries using a modified version of planimetry.
As a tool for cardiac assessment, 3D echocardiography (3DE) has largely been limited to use by experts capable of qualitative determination of left ventricular (LV) function. The usefulness of 3DE can be extended to physicians in critical care settings who have minimal training in echocardiography if it delivers quantitative parameters of LV function without expert supervision. As a critical step to generate the quantitative measures, we develop an algorithm that automatically locates and tracks the LV boundary through a sequence of 2D frames, such as those constituting a 3DE data set. A novel approach of the algorithm in this paper is the computation of an edge probability field for each frame. For unsupervised processing, the algorithm incorporates a template-based search in the first frame using prior knowledge about the LV shape. Then, by active contour (or snake) matching, the LV boundary is obtained as a MAP estimate. The internal energy constraints of the snake serve as the prior probability. Unlike the conventional snakes that depend on hard edge information as external energy, the snake in our approach uses the soft information in the edge probability field. We also can obtain a measure of confidence in the boundary estimate from the value of the energy function at the estimated contour. The estimate from one frame is used to initialize active contour matching in the next frame for LV tracking.
This paper presents an algorithm that extracts accurate left ventricular (LV) boundaries from a 2D echocardiographic (echo) sequence covering a cardiac cycle. Unlike user- dependent, manual or semi-automatic techniques, the key feature of this algorithm is its truly automated processing for estimation. First, the algorithm performs smoothing of the image in the LV target area, followed by enhancement of intensity differences and edge detection. In order to best localize the position of the LV boundary, the algorithm uses a deformable template model derived from prior knowledge of LV shape and an edge map obtained from boundary estimation. The deformable template model is matched to the target by minimizing an energy function induced by the difference between the edge locations and tangents of the template and those of the current frame edge map. Since the shape of the endocardial boundary will vary between temporally distinct frames, a controlled continuity spline, a snake, is then used to implement refined active contour matching to the current frame LV boundary. Frame-to-frame tracking of the LV boundary is incorporated by using the boundary estimate from one frame to initialize and help with the estimation in the subsequent frame, which leads to faster and more accurate LV estimation throughout the image sequence. Test results of this algorithm show that the combination of approximate template matching with smoothness constraints in snakes produces good LV boundary extraction even with significant false and/or missing edge information caused by poor contrast and noise.