Unsupervised extraction of the aortic dissection membrane based on a multiscale piecewise ridge model

This contribution expounds an unsupervised technique for successfully retrieving fine structures, such as the aortic dissection membrane, from noisy, artifact-afflicted medical image data. A model of the fine structure as a ridge-like element is employed at several scales and orientations. Towards this end, we piecewise approximate the structure's idealized intensity pattern and shape in 2D by a second derivative of a Gaussian function in cross-sectional direction and by a Gaussian in longitudinal direction. The filter responses selected at each pixel undergo a Fuzzy c-means clustering aiming at determining cluster centers and class probability distributions for discrimination between desired structures and artifacts. We remodel the obtained distribution functions in order to obtain a normalized, complete partition. With an average distance of 0.65 mm between automatic segmentation and ground truth in 12 CTA datasets of aortic dissection, our novel approach proves more accurate than previous methods.