Cross-Modality Vertebrae Localization and Labeling Using Learning-Based Approaches

Abstract

Spine is one of the major organs in human body. It consists of multiple vertebrae and inter-vertebral discs. As the locations and labels of vertebrae provide a vertical reference framework to different organs in the torso, they play an important role in various neurological, orthopaedic and oncological studies. On the other hand, however, manual localization and labeling of vertebrae is often time consuming. Therefore, automatic vertebrae localization and labeling has drawn significant attentions in the community of medical image analysis. While some pioneer studies aim to localize and label vertebrae using domain knowledge, more recent studies tackle this problem via machine learning technologies. With the spirit of “data-driven”, learning-based approaches are able to extract the appearance and geometric characteristics of vertebrae more efficient and effective than hand-crafted algorithms. More importantly, it facilitates cross-modality vertebrae localization, i.e., a generic algorithm working on different imaging modalities. In this chapter, we start with a review of several representative learning-based vertebrae localization and labeling methods. The key ideas of these methods are re-visited. In order to achieve a solution that is robust to severe diseases (e.g., scoliosis) and imaging artifacts (e.g., metal artifacts), we propose a learning-based method with two novel components. First, instead of treating vertebrae/discs as either repetitive components or completely independent entities, we emulate a radiologist and use a hierarchial strategy to learn detectors dedicated to anchor (distinctive) vertebrae, bundle (non-distinctive) vertebrae and inter-vertebral discs, respectively. At run-time, anchor vertebrae are detected concurrently to provide redundant and distributed appearance cues robust to local imaging artifacts. Bundle vertebrae detectors provide candidates of vertebrae with subtle appearance differences, whose labels are mutually determined by anchor vertebrae to gain additional robustness. Disc locations are derived from a cloud of responses from disc detectors, which is robust to sporadic voxel-level errors. Second, owing to the non-rigidness of spine anatomies, we employ a local articulated model to effectively model the spatial relations across vertebrae and discs. The local articulated model fuses appearance cues from different detectors in a way that is robust to abnormal spine geometry caused by severe diseases. Our method is validated on a large scale of CT (189) and MR (300) spine scans. It exhibits robust performance, especially to cases with severe diseases and imaging artifacts.