Image segmentation and reconstruction of 3D surfaces from carotid ultrasound images

A new algorithm is proposed for the automatic segmentation of the common carotid wall in ultrasound images. It uses the random sample consensus (RANSAC) method to estimate the most significant cubic splines or ellipses fitting the edge map of a longitudinal or a transversal section, respectively. Periodic spline fitting is also investigated for transversal sections. The geometric model (spline or ellipse) consensus is evaluated through a new gain function, which integrates the responses to different discriminating characteristics of the carotid boundary: the proximity of the geometric model to any edge pixel and to valley shaped edge pixels; the consistency between the orientation of the normal to the geometric model and the intensity gradient; and the distance to a rough estimate of the lumen boundary. To obtain an edge map with reduced noise and a good localization of the edges, a new robust non-linear smoothing filter was conceived. It preserves weaker intensity edges if they have low curvature, which is a common characteristic of anatomical structures. The smoothing filter combines the image curvature with an edge detector, known as the instantaneous coefficient of variation (ICOV), more suited to ultrasound imaging than the intensity gradient norm. Robust statistics methods are used to improve the automatic stopping of the smoothing and to produce sharper boundaries. For the detection of the lumen boundary, a new dynamic programming model was conceived for longitudinal sections and adapted to transversal sections, where the boundary is a closed curve. It looks for the path that maximizes the accumulated ICOV strength, previously normalized in the direction normal to the lumen axis. A hybrid geometric active contour is also introduced. It uses a smooth thresholding surface and the image intensity topology to reduce some segmentation errors, left by the dynamic programming algorithm, and to guide the smoothing of the detected lumen boundary. A new approach and several implementation details are investigated for the automatic reconstruction of 3D surfaces of the common carotid. The surfaces are computed from the wall and lumen boundary contours segmented in each frame of a data set acquired with a 3D ultrasound freehand system. A level set algorithm is proposed for the computation of new slices from the reconstructed 3D surfaces. A cross sectional area minimization approach is introduced to compute reliable estimates of slices normal to the longitudinal axis of the carotid artery. Statistics for the segmentation results are computed for each case, using an image data base manually segmented by a medical expert as the ground truth.