Detection of 3D Spinal Geometry Using Iterated Marginal Space Learning

Abstract

Abstract. Determining spinal geometry and in particular the position and orientation of the intervertebral disks is an integral part of nearly every spinal examination with Computed Tomography (CT) and Magnetic Resonance (MR) imaging. It is particularly important for the standardized alignment of the scan geometry with the spine. In this paper, we present a novel method that combines Marginal Space Learning (MSL), a recently introduced concept for efficient discriminative object detection, with a generative anatomical network that incorporates relative pose information for the detection of multiple objects. It is used to simultaneously detect and label the intervertebral disks in a given spinal image volume. While a novel iterative version of MSL is used to quickly generate candidate detections comprising position, orientation, and scale of the disks with high sensitivity, the anatomical network selects the most likely candidates using a learned prior on the individual nine dimensional transformation spaces. Since the proposed approach is learning-based it can be trained for MR or CT alike. Experimental results based on 42 MR volumes show that our system not only achieves superior accuracy but also is the fastest system of its kind in the literature -on average, the spinal disks of a whole spine are detected in 11.5s with 98.6% sensitivity and 7.3% false positive detections. An average position error of 2.4mm and angular error of 3.9 • is achieved.

Automatic Localization and Identification of Vertebrae in Arbitrary Field-of-View CT Scans

“…A common approach for vertebra (in CT) and intervertebral disks (in MRI) is to employ a multi-stage approach. In the first stage a detector in the form of a filter [1,2], a single/multi-class classifier [3][4][5][6][7][8] or a model-based Hough transform [9] is used to detect potential vertebra candidates. As these candidates may contain many false positive responses a second stage is applied to add robustness.…”

“…In [2], a clever search is performed based on prior information through the candidates, while [1,4,5] fit a low order polynomial curve to the candidates to remove outliers. In [3,6,8,9] the authors add prior information via graphical models, such as Hidden Markov Models (HMMs) [10], and infer the maximum a-posteriori (MAP) estimate for the vertebrae locations. In contrast, [11,12] use a fully generative model, and inference is achieved via generalized expectation-maximization, while in [7] deformable templates are used for segmentation and subsequent identification.…”

“…In fact, most algorithms require a priori knowledge of which vertebrae are visible in the scan [1][2][3][4][5][6][7][8]. They either focus on a specific region, such as lumbar or thoracic, or need to modify their models based on the expected spine region.…”

Automatic detection of lytic and blastic thoracolumbar spine metastases on computed tomography

“…1). In CT data, vertebral bodies can be reliably detected, for example, using iterated marginal space learning [13]. Vertebral bodies are highlighted on the CT image and restrict the search space for subsequent lesion detection.…”

“…The first model in the detector exploits a limited set of features which includes low-level 3D Haar-like features. The other subsequent two models exploit the full heterogeneous set of features, which are of different nature and describe various characteristics of the suspected lesion [13, 15, 16]. The lesion centre detector cascade provides rough to coarse lesion detection, which starts with a large set of suspicious lesion-like structures and ends with a reduced set of likely malignant clinically important findings.…”