Hierarchical segmentation and identification of thoracic vertebra using learning-based edge detection and coarse-to-fine deformable model.

Precise segmentation and identification of thoracic vertebrae is important for many medical imaging applications whereas it remains challenging due to vertebra's complex shape and varied neighboring structures. In this paper, a new method based on learned bone-structure edge detectors and a coarse-to-fine deformable surface model is proposed to segment and identify vertebrae in 3D CT thoracic images. In the training stage, a discriminative classifier for object-specific edge detection is trained using steerable features and statistical shape models for 12 thoracic vertebrae are also learned. In the run-time, we design a new coarse-to-fine, two-stage segmentation strategy: subregions of a vertebra first deforms together as a group; then vertebra mesh vertices in a smaller neighborhood move group-wise, to progressively drive the deformable model towards edge response maps by optimizing a probability cost function. In this manner, the smoothness and topology of vertebra's shapes are guaranteed. This algorithm performs successfully with reliable mean point-to-surface errors 0.95 +/- 0.91 mm on 40 volumes. Consequently a vertebra identification scheme is also proposed via mean surface meshes matching. We achieve a success rate of 73.1% using a single vertebra, and over 95% for 8 or more vertebra which is comparable or slightly better than state-of-the-art.