VolHOG: a volumetric object recognition approach based on bivariate histograms of oriented gradients for vertebra...

The automatic recognition of vertebrae in volumetric images is an important step toward automatic spinal diagnosis and therapy support systems. There are many applications such as the detection of pathologies and segmentation which would benefit from automatic initialization by the detection of vertebrae. One possible application is the initialization of local vertebral segmentation methods, eliminating the need for manual initialization by a human operator. Automating the initialization process would optimize the clinical workflow. However, automatic vertebra recognition in magnetic resonance (MR) images is a challenging task due to noise in images, pathological deformations of the spine, and image contrast variations.

This work presents a fully automatic algorithm for 3D cervical vertebra detection in MR images. We propose a machine learning method for cervical vertebra detection based on new features combined with a linear support vector machine for classification. An algorithm for bivariate gradient orientation histogram generation from three-dimensional raster image data is introduced which allows us to describe three-dimensional objects using the authors' proposed bivariate histograms.

A detailed performance evaluation on 21 T2-weighted MR images of the cervical vertebral region is given. A single model for cervical vertebrae C3-C7 is generated and evaluated. The results show that the generic model performs equally well for each of the cervical vertebrae C3-C7. The algorithm's performance is also evaluated on images containing various levels of artificial noise. The results indicate that the proposed algorithm achieves good results despite the presence of severe image noise.

The proposed detection method delivers accurate locations of cervical vertebrae in MR images which can be used in diagnosis and therapy. In order to achieve absolute comparability with the results of future work, the authors are following an open data approach by making the image dataset used in their performance evaluation available to the public.