Learning-based spine vertebra localization and segmentation in 3D CT image

Abstract

Spine segmentation is important for spinal screening and examination in the assistance of pathological progression evaluation and therapies. In this paper, we propose a novel solution for spine vertebra localization and segmentation in 3D volumetric CT data. Our spine vertebra localization includes: (1) spine centerline and spine canal centerline extraction and (2) vertebra centers and intervertebral disc centers localization. The final spine segmentation is based on the results of spine vertebra localization. Our solution is characterized by three key ingredients: First, we present a new and efficient way to extract spine centerline and spine canal centerline. Second, the vertebra center and intervertebral disc center localization are estimated by probabilistic inference approach. Third, a case-specific foreground and background constraints are constructed for each vertebra digit in the segmentation framework. For evaluation, the proposed method is applied to a data set which contains 10 CT volumes. Our approach achieves average detection error of 1.6 mm for both vertebra center and intervertebral disc center. Segmentation results also demonstrate the effectiveness of our method.