Multiscale segmentation of the aorta in 3D ultrasound images

Fast, reliable segmentation of the abdominal aorta from three dimensional ultrasound remains a difficult problem. Standard methods based on local information like thresholding, region growing or active contours fail in separating the arteries from the veins and suffer from the lack of homogeneity of the vessel intensity and from the partial contour information. We propose to use a model-based multiscale detection of the vessels centerlines based on a cylindrical model with circular cross-section, and adapted from previous work. Our method provides a set of centerlines and an estimate of the vessel radii along each line. After an interactive selection of the desired tines, a model of the aorta is generated using the radii information and compared to a manual segmentation. This model can also be locally improved using a level set technique in order to stick to the contours of the image and to allow non-circular cross-section.