Fully automatic shape constrained mandible segmentation from cone-beam CT data

Cone-beam CT images are useful in operative dentistry but suffer from a comparatively bad image quality with regard to the signal-to-noise ratio. Therefore, we use a statistical shape model (SSM) for robust segmentation of the mandible. In contrast to previous approaches, our method (i) is fully automatic in terms of both, the establishment of correspondence and the segmentation itself, and (ii) allows for leaving the learned principal subspace. By this means, we attain a segmentation accuracy equal to the current reference work on SSM based mandible segmentation whereas our training population is 3.5 times smaller. An important reason therefor is the establishment of correspondence by optimizing a modelbased cost function. Our results indicate that SSMs with optimized correspondence can help to improve segmentation accuracy compared to alternative approaches, thus accounting for the first time for the importance of correspondence optimization in an application for image segmentation.