Robust MR Spine Detection Using Hierarchical Learning and Local Articulated Model

Abstract

A clinically acceptable auto-spine detection system, i.e., localization and labeling of vertebrae and inter-vertebral discs, is required to have high robustness, in particular to severe diseases (e.g. scoliosis) and imaging artifacts (e.g. metal artifacts in MR). Our method aims to achieve this goal 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 (nondistinctive) 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 resulting from severe diseases. Our method is validated by 300 MR spine scout scans and exhibits robust performance, especially to cases with severe diseases and imaging artifacts.

Registration of MRI to intraoperative radiographs for target localization in spinal interventions

“…Definition of such point locations (referred to as vertebral centroids, c i ) can be done automatically in CT (Dikmen et al 2008) and MR (Zhan et al 2012). Taking these locations as seed points for the segmentation, we formulated a spatially continuous min-cut problem with the objective function: $$S=\underset{ufalse(xfalse)\in false\{0,1false\}}{\text{min}}true\int false[false(1-ufalse).D_{1}false(xfalse)+u.D_{2}false(xfalse)false]dx+true\int gfalse(xfalse)false|\nabla ufalse(xfalse)false|dx$$ where u ( x ) is a membership function defining whether each pixel x lies outside ( u ( x ) = 0) or inside ( u ( x ) = 1) the vertebrae.…”

“…Automatic labeling of vertebrae in the MR image (Zhan et al 2012, Dikmen et al 2008) could be alternatively used to improve workflow.…”

Multi-Modality Vertebra Recognition in Arbitrary Views Using 3D Deformable Hierarchical Model

“…Klinder et al [2] used generalized Hough transform model for detecting vertebrae in CT, they extracted the 3D vertebra meshes from their detection-assisted segmentation but did not consider the geometric relations between vertebra meshes. Zhan et al [3] utilized Adaboost method for learning detectors for vertebrae on MR. They proposed a hierarchical probability model of spine and applied it in inferring vertebra locations and labels, but vertebra poses and spine shape are not involved.…”

“…For vertebrae/disc labeling, [2]- [5] had very successful labeling on fully or partially scanned image volumes. The local vertebrae labels relied on the identification of some special landmarks detected from multiple image views, i.e., template models for axial view vertebrae [2], annotation of spinal canals [5] or anchor vertebrae [3] in axial views. The complete vertebrae labels in the input images are inferred by a probability inference model, i.e., a graph model [13] [14], Hidden Markov Model (HMM) [4], or hierarchical model [15] [3].…”

“…To handle the limitations of detecting/segmenting anatomical structures by local appearances, a number of hierarchical models were employed [27] [28] [29]. For segmentation, Zhan et al [3] found boundary appearance of the anatomical structure can be hierarchically modeled through an iterative global/local clustering. Ma et al [28] used the hierarchical coarse-to-fine mechanism to identify the precise edge locations of 3D CT thoracic images, providing accurate 3D segmentation.…”