Recent Advances in Computational Methods and Clinical Applications for Spine Imaging

Automated detection of sclerotic metastases (bone lesions) in Computed Tomography (CT) images has potential to be an important tool in clinical practice and research. State-of-the-art methods show performance of 79 % sensitivity or true-positive (TP) rate, at 10 false-positives (FP) per volume. We design a two-tiered coarse-to-fine cascade framework to first operate a highly sensitive candidate generation system at a maximum sensitivity of \(\sim \)92 % but with high FP level (\(\sim \)50 per patient). Regions of interest (ROI) for lesion candidates are generated in this step and function as input for the second tier. In the second tier we generate \(N\) 2D views, via scale, random translations, and rotations with respect to each ROI centroid coordinates. These random views are used to train a deep Convolutional Neural Network (CNN) classifier. In testing, the CNN is employed to assign individual probabilities for a new set of \(N\) random views that are averaged at each ROI to compute a final per-candidate classification probability. This second tier behaves as a highly selective process to reject difficult false positives while preserving high sensitivities. We validate the approach on CT images of 59 patients (49 with sclerotic metastases and 10 normal controls). The proposed method reduces the number of FP/vol. from 4 to 1.2, 7 to 3, and 12 to 9.5 when comparing a sensitivity rates of 60, 70, and 80 % respectively in testing. The Area-Under-the-Curve (AUC) is 0.834. The results show marked improvement upon previous work.

Current classification systems for adolescent idiopathic scoliosis lack information on how the spine is deformed in three dimensions (3D), which can mislead further treatment recommendations. We propose an approach to address this issue by a deep learning method for the classification of 3D spine reconstructions of patients. A low-dimensional manifold representation of the spine models was learnt by stacked auto-encoders. A K-Means++ algorithm using a probabilistic seeding method clustered the low-dimensional codes to discover sub-groups in the studied population. We evaluated the method with a case series analysis of 155 patients with Lenke Type-1 thoracic spinal deformations recruited at our institution. The clustering algorithm proposed 5 sub-groups from the thoracic population, yielding statistically significant differences in clinical geometric indices between all clusters. These results demonstrate the presence of 3D variability within a pre-defined 2D group of spinal deformities.

Accurate dose predication is critical to spinal stereotactic body radiation therapy (SBRT). It enables radiation oncologists and planners to design treatment plans that maximally protect spinal cord while effectively controlling surrounding tumors. Spinal cord dose distribution is primarily affected by the shapes of tumor boundaries near the organ. In this work, we estimate such boundary effects and predict dose distribution by exploring an active optical flow model (AOFM). To establish AOFM, we collect a sequence of dose sub-images and tumor contours near spinal cords from a database of clinically accepted spine SBRT plans. The data are classified into five groups according to the tumor location in relation to the spinal cords. In each group, we randomly choose a dose sub-image as the reference and register all other dose images to the reference using an optical flow method. AOFM is then constructed by importing optical flow vectors and dose values into the principal component analysis. To develop the predictive model for a group, we also build active shape model (ASM) of tumor contours near the spinal cords. The correlation between ASM and AOFM is estimated via the multiple regression model. When predicting dose distribution of a new case, the group was first determined based on the case’s tumor contour. Then the corresponding model for the group is used to map from the ASM space to the AOFM space. Finally, the parameters in the AOFM space are used to estimate dose distribution. This method was validated on 30 SBRT plans. Analysis of dose-volume histograms revealed that at the important 2 % volume mark, the dose difference between prediction and clinical plan is less than \(4\,\%\). These results suggest that the AOFM-based approach is a promising tool for predicting accurate spinal cord dose in clinical practice.

Monitoring spinal curvature in adolescent kyphoscoliosis requires regular radiographic examinations; however, the applied ionizing radiation increases the risk of cancer. Ultrasound imaging is favorable over X-ray because it does not emit ionizing radiation. It has been shown in the past that tracked ultrasound can be used to localize vertebral transverse processes as landmarks along the spine to measure curvature angles. Tests have been performed with spine phantoms, but scanning protocol, tracking system, data acquisition and processing time has not been considered in human subjects yet. In this paper, a portable optically tracked ultrasound system for scoliosis measurement is presented. It provides a simple way to acquire data in the clinical environment with the aim of comparing results to current X-ray-based measurement. The workflow of the procedure was tested on volunteers. The customized open-source software is shared with the community as part of our effort to make a clinically practical system.

Segmentation of the vertebrae in the spine is of relevance to many medical applications related to the spine. This paper describes a method based upon atlas-based registration for achieving an accurate segmentation of the thoracic and the lumbar vertebrae in the spine as imaged by computed tomography. The method has been evaluated on ten data sets provided as a part of the segmentation challenge hosted by the 2nd MICCAI workshop on Computational Methods and Clinical Applications for Spine Imaging. An average point-to-surface error of \(1.05\,\pm \,0.65\) mm and a mean DICE coefficient of \(0.94\,\pm \,0.03\) were obtained when comparing the computed segmentations with ground truth segmentations. These results are highly competitive when compared to the results of earlier presented methods.

We describe a fully automated approach to vertebrae segmentation from CT images which operates on superpixels. The method is based on a conditional random field model incorporating constraints learned from labelled superpixel features. The method is shown to provide consistently accurate segmentations of different vertebrae from a variety of subjects.

Detection of an object of interest can be represented as an optimization problem that can be solved by brute force or heuristic algorithms. However, the globally optimal solution may not represent the optimal detection result, which can be especially observed in the case of vertebra detection, where neighboring vertebrae are of similar appearance and shape. An adequate optimizer has to therefore consider not only the global optimum but also local optima that represent candidate locations for each vertebra. In this paper, we describe a novel framework for automated spine and vertebra detection in three-dimensional (3D) images of the lumbar spine, where we apply a novel optimization technique based on interpolation theory to detect the location of the whole spine in the 3D image and to detect the location of individual vertebrae within the spinal column. The performance of the proposed framework was evaluated on \(10\) computed tomography (CT) images of the lumbar spine. The resulting mean symmetric absolute surface distance of \(1.25\,{\pm }\,0.41\) mm and Dice coefficient of \(83.67\,{\pm }\,4.44\)%, computed from the final vertebra detection results against corresponding reference vertebra segmentations, indicate that the proposed framework can successfully detected vertebrae in CT images of the lumbar spine.

Accurate and robust segmentation of spinal and vertebral structures from medical images is a challenging task due to a relatively high degree of anatomical complexity and articulation of spinal structures, as well as due to image spatial resolution, inhomogeneity and low signal-to-noise ratio. In this paper, we describe an improved framework for vertebra segmentation that is based on an existing shape-constrained deformable model, which was modified with the aim to improve segmentation accuracy, and combined with a robust initialization that results from vertebra detection by interpolation-based optimization. The performance of the proposed segmentation framework was evaluated on \(10\) computed tomography (CT) images of the lumbar spine. The overall segmentation performance of \(0.43\,{\pm }\,0.14\) mm in terms of mean symmetric absolute surface distance and \(93.76\,{\pm }\,1.61\,\%\) in terms of Dice coefficient, computed against corresponding reference vertebra segmentations, indicates that the proposed framework can accurately segment vertebrae from CT images of the lumbar spine.

With the advances in minimal invasive surgical procedures, accurate and detailed extraction of the vertebral boundaries is required. In practice, this is a difficult challenge due to the highly complex geometry of the vertebrae, in particular at the processes. This paper presents a statistical modeling approach for detailed vertebral segmentation based on part decomposition and conditional models. To this end, a Vononoi decomposition approach is employed to ensure that each of the main subparts the vertebrae is identified in the subdivision. The obtained shape constraints are effectively relaxed, allowing for an improved encoding of the fine details and shape variability at all the regions of the vertebrae. Subsequently, in order to maintain the statistical coherence of the ensemble, conditional models are used to model the statistical inter-relationships between the different subparts. For shape reconstruction and segmentation, a robust model fitting procedure is introduced to exclude outlying inter-part relationships in the estimation of the shape parameters. The experimental results based on a database of 30 CT scans show significant improvement in accuracy with respect to the state-of-the-art and the potential of the proposed technique for detailed vertebral modeling.

Segmenting tubular structures from medical image data is a common problem; be it vessels, airways, or nervous tissue like the spinal cord. Many application-specific segmentation techniques have been proposed in the literature, but only few of them are fully automatic and even fewer approaches maintain a convex formulation. In this paper, we show how to integrate a cross-sectional similarity prior into the convex continuous max-flow framework that helps to guide segmentations in image regions suffering from noise or artefacts. Furthermore, we propose a scheme to explicitly include tubularity features in the segmentation process for increased robustness and measurement repeatability. We demonstrate the performance of our approach by automatically segmenting the cervical spinal cord in magnetic resonance images, by reconstructing its surface, and acquiring volume measurements.

This paper describes a fully automatic system for obtaining the standard Pfirrmann degeneration grading of individual intervertebral spinal discs in T2 MRI scans. It involves detecting and labeling all the vertebrae in the scan and then learning a regression from the disc region to the grading. In developing the regression function we investigate a spectrum of support regions which involve differing degrees of segmentation of the scan: our intention is to ascertain to what extent segmentation is necessary or detrimental in obtaining robust and accurate measurements. The methods are assessed on a heterogeneous clinical dataset containing 1,710 Pfirrmann-graded discs, from 285 symptomatic back pain patients. We are able to predict the grade to \(\pm 1\) precision at 85.8 % accuracy. Our novel method proposes new image features that outperform previous features and utilizes techniques to improve robustness to MR imaging variations.

This paper proposed an automated 3D lumbar intervertebral disc (IVD) segmentation strategy from MRI data. Starting from two user supplied landmarks, the geometrical parameters of all lumbar vertebral bodies and intervertebral discs are automatically extracted from a mid-sagittal slice using a graphical model based approach. After that, a three-dimensional (3D) variable-radius soft tube model of the lumbar spine column is built to guide the 3D disc segmentation. The disc segmentation is achieved as a multi-kernel diffeomorphic registration between a 3D template of the disc and the observed MRI data. Experiments on 15 patient data sets showed the robustness and the accuracy of the proposed algorithm.

A framework for 3D segmentation of healthy and herniated intervertebral discs from T2-weighted MRI was developed that exploits weak shape priors encoded in simplex mesh active surface models. An ellipsoidal simplex template mesh was initialized within the disc image boundary through affine landmark-based registration, and was allowed to deform according to image gradient forces. Coarse-to-fine multi-resolution approach was adopted in conjunction with decreasing shape memory forces to accurately capture the disc boundary. User intervention is allowed to turn off the shape feature and guide model deformation when internal shape memory influence hinders detection of pathology. For testing, 16 healthy discs were automatically segmented, and 5 pathological discs were segmented with minimal supervision. A resulting surface mesh was utilized for disc compression simulation under gravitational and weight loads and Meshless-Mechanics (MM)-based cutting using Simulation Open Framework Architecture (SOFA). The surface-mesh based segmentation method is part of a processing pipeline for anatomical modeling to support interactive surgery simulation. Segmentation results were validated against expert guided segmentation and demonstrate mean absolute shape distance error of less than 1 mm.

Fractures associated with osteoporosis are a significant public health risk, and one that is likely to increase with an ageing population. However, many osteoporotic vertebral fractures present on images do not come to clinical attention or lead to preventative treatment. Furthermore, vertebral fracture assessment (VFA) typically depends on subjective judgement by a radiologist. The potential utility of computer-aided VFA systems is therefore considerable. Previous work has shown that Active Appearance Models (AAMs) give accurate results when locating landmarks on vertebra in DXA images, but can give poor fits in a substantial subset of examples, particularly the more severe fractures. Here we evaluate Random Forest Regression Voting Constrained Local Models (RFRV-CLMs) for this task and show that, while they lead to slightly poorer median errors than AAMs, they are much more robust, reducing the proportion of fit failures by 68 %. They are thus more suitable for use in computer-aided VFA systems.

This paper proposes a parameterization method of vertebra models by mapping them onto the parameterized surface of a torus. Our method is based on a modified Self-organizing Deformable Model (mSDM) [1], which is a deformable model guided by competitive learning and an energy minimization approach. Unlike conventional mapping methods, the mSDM finds the one-to-one mapping between arbitrary surface model and the target surface with the same genus as the model. At the same time, the mSDM can preserve geometrical properties of the original model before and after mapping. Moreover, users are able to control mapping positions of the feature vertices in the model. Using the mSDM, the proposed method maps the vertebra model onto a torus surface through an intermediate surface with the approximated shape of the vertebra. The use of the intermediate surface results in the stable mapping of the vertebra to a torus compared with the direct mapping from the model to the torus.

This paper describes on-going work on the transposition to digital format of 2D images of a printed atlas of the brainstem. In MRI-based anatomical modeling for neurosurgery planning and simulation, the complexity of the functional anatomy entails a digital atlas approach, rather than less descriptive voxel or surface-based approaches. However, there is an insufficiency of descriptive digital atlases, in particular of the brainstem. Our approach proceeds from a series of numbered, contour-based sketches coinciding with slices of the brainstem featuring both closed and open contours. The closed contours coincide with functionally relevant regions, in which case our objective is to fill in each corresponding label, which is analogous to painting numbered regions in a paint-by-numbers kit. The open contours typically coincide with cranial nerve tracts as well as symbols representing the medullary pyramids. This 2D phase is needed in order to produce densely labeled regions that can be stacked to produce 3D regions, as well as identifying embedded paths and outer attachment points of cranial nerves. In future work, the stacked labeled regions will be resampled and refined probabilistically, through active contour and surface modeling based on MRI T1, T2 and tractographic data. The relevance to spine modeling of this project is two-fold: (i) this atlas will fill a void left by the spine and brain segmentation communities, as no digital atlas of the brainstem exist, and (ii) this atlas is necessary to make explicit the attachment points of major nerves having both cranial and spinal origin, specifically nerves X and XI, as well all the attachment points of cranial nerves other than I and II.

Segmentation of the vertebrae in the spine is of relevance to many medical applications. To this end, the 2nd MICCAI workshop on Computational Methods and Clinical Applications for Spine Imaging organized a segmentation challenge. This paper briefly presents one of the participating methods along with achieved results. The employed method is based upon atlas-based segmentation, where a number of atlases of the spine are registered to the target data set. The labels of the deformed atlases are combined using label fusion to obtain the final segmentation of the target data set. An average DICE score of \( 0.94 \pm 0.03 \) was achieved on the training data set.

The vertebral column is of particular importance for many clinical procedures such as anesthesia or anaelgesia. One of the main challenges for diagnostic and interventional tasks at the spine is its robust and accurate segmentation. There exist a number of segmentation approaches that mostly perform segmentation on the individual vertebrae. We present a novel segmentation approach that uses statistical multi-object shape\(+\)pose models and evaluate it on a standardized data set. We could achieve a mean dice coefficient of \(0.83\) for the segmentation. The flexibility of our approach let it become valuable for the specific segmentation challenges in clinical routine.

Automatic segmentation of 3D vertebrae is a challenging task in medical imaging. In this paper, we introduce a total variation (TV) based framework that incorporates an a priori model, i.e., a vertebral mean shape, image intensity and edge information. The algorithm was evaluated using leave-one-out cross validation on a data set containing ten computed tomography scans and ground truth segmentations provided for the CSI MICCAI 2014 spine and vertebrae segmentation challenge. We achieve promising results in terms of the Dice Similarity Coefficient (DSC) of \(0.93 \pm 0.04\) averaged over the whole data set.

This paper presents a method for automatic vertebra segmentation. The method consists of two parts: vertebra detection and vertebra segmentation. To detect vertebrae in an unknown CT spine image, an interpolation-based optimization approach is first applied to detect the whole spine, then to detect the location of individual vertebrae, and finally to rigidly align shape models of individual vertebrae to the detected vertebrae. Each optimization is performed using a spline-based interpolation function on an equidistant sparse optimization grid to obtain the optimal combination of translation, scaling and/or rotation parameters. The computational complexity in examining the parameter space is reduced by a dimension-wise algorithm that iteratively takes into account only a subset of parameter space dimensions at the time. The obtained vertebra detection results represent a robust and accurate initialization for the subsequent segmentation of individual vertebrae, which is built upon the existing shape-constrained deformable model approach. The proposed iterative segmentation consists of two steps that are executed in each iteration. To find adequate boundaries that are distinctive for the observed vertebra, the boundary detection step applies an improved robust and accurate boundary detection using Canny edge operator and random forest regression model that incorporates prior knowledge through image intensities and intensity gradients. The mesh deformation step attracts the mesh of the vertebra shape model to vertebra boundaries and penalizes the deviations of the mesh from the training repository while preserving shape topology.

In this paper, a former method has been adapted to perform vertebra segmentations for the 2nd Workshop on Computational Methods and Clinical Applications for Spine Imaging (CSI 2014). A statistical Shape Models (SSM) of each lumbar vertebra was created for the segmentation step. From manually placed intervertebral discs centres, the similarity parameters are computed to initialise the vertebra shapes. The segmentation is performed by iteratively deforming a mesh inside the image intensity and then projecting it into the SSM space until convergence. Afterwards, a relaxation step based on B-spline is applied to overcome the SSM rigidity. The deformation of the mesh, within the image intensity, is performed by displacing each landmark along the normal direction of the surface mesh at the landmark position seeking a minimum of a cost function based on a set of trained features. The organisers tested the performance of our method with a dataset of five patients, achieving a global mean Dice Similarity Index (DSI) of 93.4 %. Results were consistent and accurate along the lumbar spine 93.8, 93.9, 93.7, 93.4 and 92.1 %, from L1 to L5.

Segmentation is the fundamental step for most spine image analysis tasks. The vertebra segmentation challenge held at the 2014 Computational Spine Imaging Workshop (CSI2014) objectively evaluated the performance of several algorithms segmenting vertebrae in spine CT scans. Five teams participated in the challenge. Ten training data sets and Five test data sets with reference annotation were provided for training and evaluation. Dice coefficient and absolute surface distances were used as the evaluation metrics. The segmentations on both the whole vertebra and its substructures were evaluated. The performances comparisons were assessed in different aspects. The top performers in the challenge achieved Dice coefficient of 0.93 in the upper thoracic, 0.95 in the lower thoracic and 0.96 in the lumbar spine. The strength and weakness of each method are discussed in this paper.