Vertebral shape: automatic measurement with dynamically sequenced active appearance models

Automatic segmentation of vertebrae in 3D CT images using adaptive fast 3D pulse coupled neural networks

Two systems are presented for segmentation of vertebrae in a 3D computed tomography (CT) image. The first method extracts seven features from each voxel and uses a multi-layer perceptron neural network (MLPNN) to classify the voxel as vertebrae or background. In the second method, the segmentation is completed in two steps: first, a newly developed adaptive pulse coupled neural network (APCNN) directly applied to a given image segments vertebrae, then the result is refined using a median filter. In the developed APCNN, the values for the user-defined parameters of the pulse coupled neural networks (PCNN) are adaptively adjusted for each image individually, instead of using one value for all images as in conventional PCNN. The performance of both systems in terms of Dice index (DI) was evaluated and compared against the state-of-the-art segmentation methods using seventeen clinical and standard CT images. Overall, both systems demonstrated statistically similar and promising performance with average DI > 95%. Compared to existing PCNN-based segmentation algorithms, the accuracy of the proposed APCNN improved by 29.3% on average. The developed APCNN-based system is more accurate than MLPNN-based system and existing PCNN-based algorithms in segmentation of vertebrae with blurred and weak boundaries and in the images contaminated by salt- and- pepper noise. In terms of computation time, the APCNN-based system is 16 times faster than the MLPNN-based system. Consequently, the presented APCNN-based algorithm is both accurate and fast and could be used in clinical environment for segmentation of vertebrae in 3D CT images.

Identification of prevalent vertebral fractures using CT lateral scout views: a comparison of semi-automated quantitative vertebral morphometry and radiologist semi-quantitative grading

We compared vertebral fracture assessment by semi-automated quantitative vertebral morphometry measurements with the conventional semi-quantitative (SQ) grading using lateral CT scout views. The semi-automated morphometry method showed good to excellent agreement with the visual SQ grading by radiologists for identification of vertebral fractures.

INTRODUCTION

Semi-automated quantitative vertebral morphometry (QM) measurements may enhance management of osteoporosis patients by providing an efficient means to identify vertebral fractures (VFx). We compared identification of prevalent VFx by semi-automated QM to SQ grading.

METHODS

A non-radiologist performed semi-automated QM from CT lateral scout views in 200 subjects (102 men, 98 women, 65.8 ± 8.9 years) selected from the Framingham Heart Study Multidetector CT Study. VFx were classified in the QM approach based on using Genant's criteria for deformities, and compared with conventional SQ grading performed by experienced radiologists as the gold standard. The kappa (k) statistics, percent agreement (% Agree), sensitivity (SE), specificity (SP), positive predictive value (PPV), and negative predictive value (NPV) were computed.

RESULTS

Among 200 subjects, 57 had mild and 41 had moderate or severe VFx by visual SQ grading. Per-person analyses showed excellent agreement between the two methods, with k = 0.780. The % Agree ranged from 86.7% to 91.2%, the SE was 81.3%-96%, and the SP was 86.5%-92%. Among 2,588 vertebrae analyzed, 107 had mild and 49 had moderate or severe VFx by visual SQ grading. Per-vertebra analyses revealed good agreement, with k = 0.580. Agreement between the methods tended to be highest in L1-L4 region. Agreement and validity measures were higher when only moderate and severe fractures were included.

CONCLUSION

The semi-automated quantitative vertebral morphometry measurements from CT lateral scout views provided good to excellent agreement with the standard SQ grading for assessment of prevalent vertebral fractures.

A 4D statistical model of wrist bone motion patterns

Direct imaging of ligament damage in the wrist remains a challenge. Still, such damage can be assessed indirectly through the analysis of changes in wrist pose and motion pattern. For this purpose we built a statistical reference model that describes healthy motion patterns. We show that such a model can also be used to detect and quantify pathologies. A model that only describes the global translations and rotations of the carpal bones is insufficiently accurate due to size and shape variations of the bones. We present a local statistical motion model that minimizes the influence of size and shape differences by analyzing the coordinate differences of pairs of points on adjacent bone surfaces. These differences are determined in a set of 14 healthy example wrists imaged in a range of poses by means of 4D-RX imaging. The distribution of the differences as a function of the pose form the local statistical motion model (LSMM). Translations of 2 mm and rotations of 20° with respect to the healthy example wrists are detected as outliers in the point pair distributions. An evaluation involving wrists with a damaged ligament between scaphoid and lunate shows that not only joint space widenings can be detected, but also shifts of congruent bone surfaces. The LSMM is also used to perform a virtual reconstruction of the most likely healthy wrist after a simulated perturbation of bones. The reconstruction precision is shown to be about 1 mm. Therefore, the presented 4D statistical model of wrist bone movement may become a valuable clinical tool for diagnosis and surgical planning.

Measurement error in statistical models of shape

Active shape models (ASMs) are popular and sophisticated methods of extracting features in (especially medical) images. Here we analyse the error in placing ASM points on the boundary of the feature. By using replications, a corrected covariance matrix is presented that should reduce the effects of placement error. We show analytically and via simulations that the cumulative variability for a given number of eigenvalues retained in principal components analysis (PCA) ought to be reduced by increasing levels of point-placement error. Results for predicted errors are in excellent agreement with the set-up parameters of two simulated shapes and with anecdotal evidence from the trained experts for real data taken from the OSTEODENT project. We derive an equation for the reliability of placing the points and we find values of 0.79 and 0.85 (where 0=bad and 1=good) for the two clinical experts for the OSTEODENT data. These analyses help us to understand the sources and effects of measurement error in shape models.

Heterogeneous computing for vertebra detection and segmentation in x-ray images

The context of this work is related to the vertebra segmentation. The method we propose is based on the active shape model (ASM). An original approach taking advantage of the edge polygonal approximation was developed to locate the vertebra positions in a X-ray image. Despite the fact that segmentation results show good efficiency, the time is a key variable that has always to be optimized in a medical context. Therefore, we present how vertebra extraction can efficiently be performed in exploiting the full computing power of parallel (GPU) and heterogeneous (multi-CPU/multi-GPU) architectures. We propose a parallel hybrid implementation of the most intensive steps enabling to boost performance. Experimentations have been conducted using a set of high-resolution X-ray medical images, showing a global speedup ranging from 3 to 22, by comparison with the CPU implementation. Data transfer times between CPU and GPU memories were included in the execution times of our proposed implementation.

Cervical vertebrae tracking in video-fluoroscopy using the normalized gradient field

For patients with neck problems valuable functional and diagnostic information can be obtained from a fluoroscopy video of a flexion-extension movement of the cervical spine. In most cases physicians have to manually extract the vertebrae, making the analysis of these video sequences tedious and time consuming. In this paper we propose an automatic fast and precise method for tracking cervical vertebrae. Our method relies only on a rough selection of template areas of each vertebra in a single frame of the video sequence. Compared to existing automated methods, no contours need to be extracted and no vertebra segmentation is required. Tracking is done with a normalized gradient field, using only the gradient orientations as features. Experimental results show that the algorithm is robust and able to track the vertebrae accurately even if they are partially occluded or if a disc prosthesis is present.

Robust autonomous model learning from 2D and 3D data sets

In this paper we propose a weakly supervised learning algorithm for appearance models based on the minimum description length (MDL) principle. From a set of training images or volumes depicting examples of an anatomical structure, correspondences for a set of landmarks are established by group-wise registration. The approach does not require any annotation. In contrast to existing methods no assumptions about the topology of the data are made, and the topology can change throughout the data set. Instead of a continuous representation of the volumes or images, only sparse finite sets of interest points are used to represent the examples during optimization. This enables the algorithm to efficiently use distinctive points, and to handle texture variations robustly. In contrast to standard elasticity based deformation constraints the MDL criterion accounts for systematic deformations typical for training sets stemming from medical image data. Experimental results are reported for five different 2D and 3D data sets.

Quantitative Vertebral Morphometry Using Neighbor-Conditional Shape Models

A novel method for vertebral fracture quantification from X-ray images is presented. Using pairwise conditional shape models trained on a set of healthy spines, the most likely normal vertebra shapes are estimated conditional on all other vertebrae in the image. The differences between the true shape and the reconstructed normal shape is subsequently used as a measure of abnormality. In contrast with the current (semi-)quantitative grading strategies this method takes the full shape into account, it uses a patient-specific reference by combining population-based information on biological variation in vertebra shape and vertebra interrelations, and it provides a continuous measure of deformity. The method is demonstrated on 212 lateral spine radiographs with in total 78 fractures. The distance between prediction and true shape is 1.0 mm for unfractured vertebrae and 3.7 mm for fractures, which makes it possible to diagnose and assess the severity of a fracture.