Quantitative diagnosis of 3D scoliotic deformities depends on a number of dedicated measurements. Existing methods rely on the manual determination of a series of anatomical landmarks in X-ray images. We have developed an automatic method to alleviate the burden of this tedious task. Our method looks for a compromise between local image information and global prior constraints and finds the most probable points using dynamic programming optimization. Remaining errors can be quickly corrected by effective user interaction. The first results are promising.