This paper presents a new automatic technique for the segmentation of the rib cage on CT images. Motivated by a usage scenario in the context of large, heterogeneous databases of CT-images, we introduce two shape descriptors to be used in conjunction with a Random Forests (RF) classifier. These descriptors were specifically designed to address the challenges of rib identification under various acquisition conditions affecting subject’s orientation and image quality. Extensive experiments demonstrate the superiority of our proposed shape descriptors in nominal configurations. Robustness with respect to subject’s orientation variation and additive noise is also demonstrated, with an improvement of classification performance of up to 25%, comparing to intensity-based descriptors, without neither pre-registration nor pre-smoothing.