Contour definition and tracking in cardiac imaging through the integration of knowledge and image evidence

The main contribution of this paper is the use of simple processing techniques, incorporated in a new multistage approach, to automatically delineate left ventricle contours. Another contribution is the proposal of the centerline distances for contour comparison, which promises a more accurate measurement than the common method, based on the distance to the closest point. Edges are detected by Gaussian filtering at coarse and fine scale. The region of interest is defined as a binary map where coarse edges are extracted throughout image sequence. A contour template is matched against the gradient of the first image. Candidate boundary points are instantiated by scanning the coarse edge map perpendicularly to the matched template. A candidate contour is estimated from these points by maximizing an edge likelihood function. A region growing algorithm gives another candidate contour. Both edge and region candidate contours are then integrated with the edge map computed at fine scale by maximizing another likelihood function. Evaluation was carried out on 12 echocardiographic and 4 angiocardiographic sequences (for a total of 289 frames). Distances between computer-generated contours and the contours traced by three experts were within interobserver variability, unlike the results obtained by Acoustic Quantification and by a general-purpose deformable model.