Thanks for the clarification. If I understand correctly, the features you seek should be extracted from an open-curve that represents dorsal portion of the interververtebral disk space. In this regard, I can think of a couple curvature-based features that can be used to discriminate between 'normal' and 'abnormal' cases. The first, and simplest, feature would be mean curvature of the curve (obtained by integrating curvature along the curve and dividing by the total arc length). When using this feature, one would expect that pathological cases would have higher mean curvature. Second feature would be a histogram of curvatures. This is a generalization of the first feature. One would expect that histograms of pathological samples would have thicker tails and modes shifted to the right. In any case, going forward, it would helpful if you posted some training data on here, so that I, and other people on here have something to work with.