Heart rate estimation from ballistocardiographic signals using deep learning.

Objective.Ballistocardiography (BCG) is an unobtrusive approach for cost-effective and patient-friendly health monitoring. In this work, deep learning methods are used for heart rate estimation from BCG signals and are compared against five digital signal processing methods found in literature.Approach.The models are evaluated on a dataset featuring BCG recordings from 42 patients, acquired with a pneumatic system. Several different deep learning architectures, including convolutional, recurrent and a combination of both are investigated. Besides model performance, we are also concerned about model size and specifically investigate less complex and smaller networks.Main results.Deep learning models outperform traditional methods by a large margin. Across 14 patients in a held-out testing set, an architecture with stacked convolutional and recurrent layers achieves an average mean absolute error (MAE) of 2.07 beat min-1, whereas the best-performing traditional method reaches 4.24 beat min-1. Besides smaller errors, deep learning models show more consistent performance across different patients, indicating the ability to better deal with inter-patient variability, a prevalent issue in BCG analysis. In addition, we develop a smaller version of the best-performing architecture, that only features 8283 parameters, yet still achieves an average MAE of 2.32 beat min-1on the testing set.Significance.This is the first study that applies and compares different deep learning architectures to heart rate estimation from bed-based BCG signals. Compared to signal processing algorithms, deep learning models show dramatically smaller errors and more consistent results across different individuals. The results show that using smaller models instead of excessively large ones can lead to sufficient performance for specific biosignal processing applications. Additionally, we investigate the use of fully convolutional networks for 1D signal processing, which is rarely applied in literature.

Assessment of the individual fracture risk of the proximal femur by using statistical appearance models.

PURPOSE: Standard diagnostic techniques to quantify bone mineral density (BMD) include dual-energy x-ray absorptiometry (DXA) and quantitative computed tomography. However, BMD alone is not sufficient to predict the fracture risk for an individual patient. Therefore, the development of tools, which can assess the bone quality in order to predict individual biomechanics of a bone, would mean a significant improvement for the prevention of fragility fractures. In this study, a new approach to predict the fracture risk of proximal femora using a statistical appearance model will be presented. METHODS: 100 CT data sets of human femur cadaver specimens are used to create statistical appearance models for the prediction of the individual fracture load (FL). Calculating these models offers the possibility to use information about the inner structure of the proximal femur, as well as geometric properties of the femoral bone for FL prediction. By applying principal component analysis, statistical models have been calculated in different regions of interest. For each of these models, the individual model parameters for each single data set were calculated and used as predictor variables in a multilinear regression model. By this means, the best working region of interest for the prediction of FL was identified. The accuracy of the FL prediction was evaluated by using a leave-one-out cross validation scheme. Performance of DXA in predicting FL was used as a standard of comparison. RESULTS: The results of the evaluative tests demonstrate that significantly better results for FL prediction can be achieved by using the proposed model-based approach (R = 0.91) than using DXA-BMD (R = 0.81) for the prediction of fracture load. CONCLUSIONS: The results of the evaluation show that the presented model-based approach is very promising and also comparable to studies that partly used higher image resolutions for bone quality assessment and fracture risk prediction.

Texture analysis, bone mineral density, and cortical thickness of the proximal femur: fracture risk prediction.

OBJECTIVE: The objectives of this study were to perform a clinical study analyzing bone quality in multidetector computed tomographic images of the femur using bone mineral density (BMD), cortical thickness, and texture algorithms in differentiating osteoporotic fracture and control subjects; to differentiate fracture types. METHODS: Femoral head, trochanteric, intertrochanteric, and upper and lower neck were segmented (fracture, n = 30; control, n = 10). Cortical thickness, BMD, and texture analysis were obtained using co-occurrence matrices, Minkowski dimension, and functional and scaling index method. RESULTS: Bone mineral density and cortical thickness performed best in the neck region, and texture measures performed best in the trochanter. Only cortical thickness and texture measures differentiated femoral neck and intertrochanteric fractures. CONCLUSIONS: This study demonstrates that differentiation of osteoporotic fracture subjects and controls is achieved with texture measures, cortical thickness, and BMD; however, performance is region specific.

Training of head and neck segmentation networks with shape prior on small datasets.

PURPOSE: Cancer in the head and neck area is commonly treated with radiotherapy. A key step for low-risk treatment is the accurate delineation of organs at risk in the planning imagery. The success of deep learning in image segmentation led to automated algorithms achieving human expert performance on certain datasets. However, such algorithms require large datasets for training and fail to segment previously unseen pathologies, where human experts still succeed. As pathologies are rare and large datasets costly to generate, we investigate the effect of: reduced training data, batch sizes and incorporation of prior knowledge. METHODS: The small data problem is studied by training a full-volume segmentation network with the reduced amount of data from the MICCAI 2015 head and neck segmentation challenge. To improve the segmentation, we evaluate the batch size as a hyper-parameter and first study and then incorporate a stacked autoencoder as shape prior into the training process. RESULTS: We found that using half of the training data (12 images of 25) results in an accuracy drop of only 3% for the segmentation of organs at risk. Also, the batch size turns out to be relevant for the quality of the segmentation when trained with less than half of the data. By applying PCA on the autoencoder's latent space we achieve a compact and accurate shape model, which is used as a regularizer and significantly improves the segmentation results. CONCLUSION: Small training data of up to 12 training images is enough to train accurate head and neck segmentation models. By using a shape prior for regularization, the performance of the segmentation can be improved significantly on the full dataset. When training on fewer than 12 images, the batch size is relevant and models have to be trained much longer until convergence.

Ejection Wave Segmentation for Contact-Free Heart Rate Estimation from Ballistocardiographic Signals.

We present a new algorithm for peak detection in ballistocardiographic (BCG) signals and use it for heart rate estimation. Systolic complexes of the BCG signal are enhanced and coarse heart beat locations estimated. Ejection waves I, J and K are detected simultaneously around coarse locations, only using detection of local maxima and weighted summation of peak heights. Due to a lack of reference BCG annotations, the algorithm's performance is assessed by using the detected peaks for heart rate estimation. On a dataset acquired with a pneumatic BCG system, we evaluate the heart rate estimation performance and compare the introduced algorithm against other methods found in literature. The dataset is gathered from 42 patients in a clinical environment and provides low-quality signals taken from a realistic scenario. With a mean absolute percentage error of 2.58 % at 65 % coverage, the presented method is on par with the best-performing state-of-the-art algorithms investigated. Limits of agreement (5th/95th percentiles) in a comparison with ECG-based heart rate measurements lie within P5 = -3.63 and P95 = 5.78 beat/min.

Multi-organ segmentation of the head and neck area: an efficient hierarchical neural networks approach.

PURPOSE: In radiation therapy, a key step for a successful cancer treatment is image-based treatment planning. One objective of the planning phase is the fast and accurate segmentation of organs at risk and target structures from medical images. However, manual delineation of organs, which is still the gold standard in many clinical environments, is time-consuming and prone to inter-observer variations. Consequently, many automated segmentation methods have been developed. METHODS: In this work, we train two hierarchical 3D neural networks to segment multiple organs at risk in the head and neck area. First, we train a coarse network on size-reduced medical images to locate the organs of interest. Second, a subsequent fine network on full-resolution images is trained for a final accurate segmentation. The proposed method is purely deep learning based; accordingly, no pre-registration or post-processing is required. RESULTS: The approach has been applied on a publicly available computed tomography dataset, created for the MICCAI 2015 Auto-Segmentation challenge. In an extensive evaluation process, the best configurations for the trained networks have been determined. Compared to the existing methods, the presented approach shows state-of-the-art performance for the segmentation of seven different structures in the head and neck area. CONCLUSION: We conclude that 3D neural networks outperform the most existing model- and atlas-based methods for the segmentation of organs at risk in the head and neck area. The ease of use, high accuracy and the test time efficiency of the method make it promising for image-based treatment planning in clinical practice.

Model-Based Vestibular Afferent Stimulation: Evaluating Selective Electrode Locations and Stimulation Waveform Shapes.

A dysfunctional vestibular system can be a severe detriment to the quality of life of a patient. Recent studies have shown the feasibility for a vestibular implant to restore rotational sensation via electrical stimulation of vestibular ampullary nerves. However, the optimal stimulation site for selective elicitation of the desired nerve is still unknown. We realized a finite element model on the basis of µCT scans of a human inner ear and incorporated naturally distributed, artificial neural trajectories. A well-validated neuron model of myelinated fibers was incorporated to predict nerve responses to electrical stimulation. Several virtual electrodes were placed in locations of interest inside the bony labyrinth (intra-labyrinthine) and inside the temporal bone, near the target nerves (extra-labyrinthine), to determine preferred stimulation sites and electrode insertion depths. We investigated various monopolar and bipolar electrode configurations as well as different pulse waveform shapes for their ability to selectively stimulate the target nerve and for their energy consumption. The selectivity was evaluated with an objective measure of the fiber recruitment. Considerable differences of required energy and achievable selectivity between the configurations were observed. Bipolar, intra-labyrinthine electrodes provided the best selectivities but also consumed the highest amount of energy. Bipolar, extra-labyrinthine configurations did not offer any advantages compared to the monopolar approach. No selective stimulation could be performed with the monopolar, intra-labyrinthine approach. The monopolar, extra-labyrinthine electrodes required the least energy for satisfactory selectivities, making it the most promising approach for functional vestibular implants. Different pulse waveform shapes did not affect the achieved selectivity considerably but shorter pulse durations showed consistently a more selective activation of the target nerves. A cathodic, centered triangular waveform shape was identified as the most energy-efficient of the tested shapes. Based on these simulations we are able to recommend the monopolar, extra-labyrinthine stimulation approach with cathodic, centered triangular pulses as good trade-off between selectivity and energy consumption. Future implant designs could benefit from the findings presented here.

Analysis of Vestibular Labyrinthine Geometry and Variation in the Human Temporal Bone.

Stable posture and body movement in humans is dictated by the precise functioning of the ampulla organs in the semi-circular canals. Statistical analysis of the interrelationship between bony and membranous compartments within the semi-circular canals is dependent on the visualization of soft tissue structures. Thirty-one human inner ears were prepared, post-fixed with osmium tetroxide and decalcified for soft tissue contrast enhancement. High resolution X-ray microtomography images at 15 µm voxel-size were manually segmented. This data served as templates for centerline generation and cross-sectional area extraction. Our estimates demonstrate the variability of individual specimens from averaged centerlines of both bony and membranous labyrinth. Centerline lengths and cross-sectional areas along these lines were identified from segmented data. Using centerlines weighted by the inverse squares of the cross-sectional areas, plane angles could be quantified. The fit planes indicate that the bony labyrinth resembles a Cartesian coordinate system more closely than the membranous labyrinth. A widening in the membranous labyrinth of the lateral semi-circular canal was observed in some of the specimens. Likewise, the cross-sectional areas in the perilymphatic spaces of the lateral canal differed from the other canals. For the first time we could precisely describe the geometry of the human membranous labyrinth based on a large sample size. Awareness of the variations in the canal geometry of the membranous and bony labyrinth would be a helpful reference in designing electrodes for future vestibular prosthesis and simulating fluid dynamics more precisely.

Multi atlas based segmentation: should we prefer the best atlas group over the group of best atlases?

Multi atlas based segmentation (MABS) uses a database of atlas images, and an atlas selection process is used to choose an atlas subset for registration and voting. In the current state of the art, atlases are chosen according to a similarity criterion between the target subject and each atlas in the database. In this paper, we propose a new concept for atlas selection that relies on selecting the best performing group of atlases rather than the group of highest scoring individual atlases. Experiments were performed using CT images of 50 patients, with contours of brainstem and parotid glands. The dataset was randomly split into two groups: 20 volumes were used as an atlas database and 30 served as target subjects for testing. Classic oracle selection, where atlases are chosen by the highest dice similarity coefficient (DSC) with the target, was performed. This was compared to oracle group selection, where all the combinations of atlas subgroups were considered and scored by computing DSC with the target subject. Subsequently, convolutional neural networks were designed to predict the best group of atlases. The results were also compared with the selection strategy based on normalized mutual information (NMI). Oracle group was proven to be significantly better than classic oracle selection (p < 10-5). Atlas group selection led to a median ± interquartile DSC of 0.740 ± 0.084, 0.718 ± 0.086 and 0.670 ± 0.097 for brainstem and left/right parotid glands respectively, outperforming NMI selection 0.676 ± 0.113, 0.632 ± 0.104 and 0.606 ± 0.118 (p < 0.001) as well as classic oracle selection. The implemented methodology is a proof of principle that selecting the atlases by considering the performance of the entire group of atlases instead of each single atlas leads to higher segmentation accuracy, being even better then current oracle strategy. This finding opens a new discussion about the most appropriate atlas selection criterion for MABS.

Nine-degrees-of-freedom flexmap for a cone-beam computed tomography imaging device with independently movable source and detector.

PURPOSE: Couch-mounted cone-beam computed tomography (CBCT) imaging devices with independently rotatable x-ray source and flat-panel detector arms for acquisitions of arbitrary regions of interest (ROI) have recently been introduced in image-guided radiotherapy (IGRT). This work analyzes mechanical limitations and gravity-induced effects influencing the geometric accuracy of images acquired with arbitrary angular constellations of source and detector in nonisocentric trajectories, which is considered essential for IGRT. In order to compensate for geometric inaccuracies of this modality, a 9-degrees-of-freedom (9-DOF) flexmap correction approach is presented, focusing especially on the separability of the flexmap parameters of the independently movable components of the device. METHODS: The 9-DOF comprise a 3D translation of the x-ray source focal spot, a 3D translation of the flat-panel's active area center and three Euler-rotations of the detector's row and column vectors. The flexmap parameters are expressed with respect to the angular position of each of the devices arms. Estimation of the parameters is performed, using a CT-based structure set of a table-mounted, cylindrical ball-bearing phantom. Digitally reconstructed radiograph (DRR) patches are derived from the structure set followed by local 2D in-plane registration and subsequent 3D transform estimation by nonlinear regression with outlier detection. RESULTS: Flexmap parameter evaluations for the factory-calibrated system in clockwise and counter-clockwise rotation direction have shown only minor differences for the overall set of flexmap parameters. High short-term reproducibility of the flexmap parameters has been confirmed by experiments over 10 acquisitions for both directions, resulting in standard deviation values of ≤0.183 mm for translational components and ≤0.0219 deg for rotational components, respectively. A comparison of isocentric and nonisocentric flexmap evaluations showed that the mean differences of the parameter curves reside within their standard deviations, confirming the ability of the proposed calibration method to handle both types of trajectories equally well. Reconstructions of 0.1 mm and 0.25 mm steel wires showed similar results for the isocentric and nonisocentric cases. The full-width at half maximum (FWHM) measure indicates an average improvement of the calibrated reconstruction of 85% over the uncalibrated reconstruction. The contrast of the point spread function (PSF) improved by 310% on average over all experiments. Moreover, a reduced amount of artifacts visible in nonisocentric reconstructions of a head phantom and a line-pair phantom has been achieved by separate application of the 9-DOF flexmap on the geometry described by the independently moving source arm and detector arm. CONCLUSIONS: Using a 9-DOF flexmap approach for correcting the geometry of projections acquired with a device capable of independent movements of the source and panel arms has been shown to be essential for IGRT use cases such as CBCT reconstruction and 2D/3D registration tasks. The proposed pipeline is able to create flexmap curves which are easy to interpret, useful for mechanical description of the device and repetitive quality assurance as well as system-level preventive maintenance. Application of the flexmap has shown improvements of image quality for planar imaging and volumetric imaging which is crucial for patient alignment accuracy.

Model-based Vestibular Afferent Stimulation: Modular Workflow for Analyzing Stimulation Scenarios in Patient Specific and Statistical Vestibular Anatomy.

Our sense of balance and spatial orientation strongly depends on the correct functionality of our vestibular system. Vestibular dysfunction can lead to blurred vision and impaired balance and spatial orientation, causing a significant decrease in quality of life. Recent studies have shown that vestibular implants offer a possible treatment for patients with vestibular dysfunction. The close proximity of the vestibular nerve bundles, the facial nerve and the cochlear nerve poses a major challenge to targeted stimulation of the vestibular system. Modeling the electrical stimulation of the vestibular system allows for an efficient analysis of stimulation scenarios previous to time and cost intensive in vivo experiments. Current models are based on animal data or CAD models of human anatomy. In this work, a (semi-)automatic modular workflow is presented for the stepwise transformation of segmented vestibular anatomy data of human vestibular specimens to an electrical model and subsequently analyzed. The steps of this workflow include (i) the transformation of labeled datasets to a tetrahedra mesh, (ii) nerve fiber anisotropy and fiber computation as a basis for neuron models, (iii) inclusion of arbitrary electrode designs, (iv) simulation of quasistationary potential distributions, and (v) analysis of stimulus waveforms on the stimulation outcome. Results obtained by the workflow based on human datasets and the average shape of a statistical model revealed a high qualitative agreement and a quantitatively comparable range compared to data from literature, respectively. Based on our workflow, a detailed analysis of intra- and extra-labyrinthine electrode configurations with various stimulation waveforms and electrode designs can be performed on patient specific anatomy, making this framework a valuable tool for current optimization questions concerning vestibular implants in humans.

Evaluation of segmentation methods on head and neck CT: Auto-segmentation challenge 2015.

PURPOSE: Automated delineation of structures and organs is a key step in medical imaging. However, due to the large number and diversity of structures and the large variety of segmentation algorithms, a consensus is lacking as to which automated segmentation method works best for certain applications. Segmentation challenges are a good approach for unbiased evaluation and comparison of segmentation algorithms. METHODS: In this work, we describe and present the results of the Head and Neck Auto-Segmentation Challenge 2015, a satellite event at the Medical Image Computing and Computer Assisted Interventions (MICCAI) 2015 conference. Six teams participated in a challenge to segment nine structures in the head and neck region of CT images: brainstem, mandible, chiasm, bilateral optic nerves, bilateral parotid glands, and bilateral submandibular glands. RESULTS: This paper presents the quantitative results of this challenge using multiple established error metrics and a well-defined ranking system. The strengths and weaknesses of the different auto-segmentation approaches are analyzed and discussed. CONCLUSIONS: The Head and Neck Auto-Segmentation Challenge 2015 was a good opportunity to assess the current state-of-the-art in segmentation of organs at risk for radiotherapy treatment. Participating teams had the possibility to compare their approaches to other methods under unbiased and standardized circumstances. The results demonstrate a clear tendency toward more general purpose and fewer structure-specific segmentation algorithms.

Technical Note: plastimatch mabs, an open source tool for automatic image segmentation.

PURPOSE: Multiatlas based segmentation is largely used in many clinical and research applications. Due to its good performances, it has recently been included in some commercial platforms for radiotherapy planning and surgery guidance. Anyway, to date, a software with no restrictions about the anatomical district and image modality is still missing. In this paper we introduce plastimatch mabs, an open source software that can be used with any image modality for automatic segmentation. METHODS: plastimatch mabs workflow consists of two main parts: (1) an offline phase, where optimal registration and voting parameters are tuned and (2) an online phase, where a new patient is labeled from scratch by using the same parameters as identified in the former phase. Several registration strategies, as well as different voting criteria can be selected. A flexible atlas selection scheme is also available. To prove the effectiveness of the proposed software across anatomical districts and image modalities, it was tested on two very different scenarios: head and neck (H&N) CT segmentation for radiotherapy application, and magnetic resonance image brain labeling for neuroscience investigation. RESULTS: For the neurological study, minimum dice was equal to 0.76 (investigated structures: left and right caudate, putamen, thalamus, and hippocampus). For head and neck case, minimum dice was 0.42 for the most challenging structures (optic nerves and submandibular glands) and 0.62 for the other ones (mandible, brainstem, and parotid glands). Time required to obtain the labels was compatible with a real clinical workflow (35 and 120 min). CONCLUSIONS: The proposed software fills a gap in the multiatlas based segmentation field, since all currently available tools (both for commercial and for research purposes) are restricted to a well specified application. Furthermore, it can be adopted as a platform for exploring MABS parameters and as a reference implementation for comparing against other segmentation algorithms.

Contour-Driven Atlas-Based Segmentation.

We propose new methods for automatic segmentation of images based on an atlas of manually labeled scans and contours in the image. First, we introduce a Bayesian framework for creating initial label maps from manually annotated training images. Within this framework, we model various registration- and patch-based segmentation techniques by changing the deformation field prior. Second, we perform contour-driven regression on the created label maps to refine the segmentation. Image contours and image parcellations give rise to non-stationary kernel functions that model the relationship between image locations. Setting the kernel to the covariance function in a Gaussian process establishes a distribution over label maps supported by image structures. Maximum a posteriori estimation of the distribution over label maps conditioned on the outcome of the atlas-based segmentation yields the refined segmentation. We evaluate the segmentation in two clinical applications: the segmentation of parotid glands in head and neck CT scans and the segmentation of the left atrium in cardiac MR angiography images.

Vision 20/20: perspectives on automated image segmentation for radiotherapy.

Due to rapid advances in radiation therapy (RT), especially image guidance and treatment adaptation, a fast and accurate segmentation of medical images is a very important part of the treatment. Manual delineation of target volumes and organs at risk is still the standard routine for most clinics, even though it is time consuming and prone to intra- and interobserver variations. Automated segmentation methods seek to reduce delineation workload and unify the organ boundary definition. In this paper, the authors review the current autosegmentation methods particularly relevant for applications in RT. The authors outline the methods' strengths and limitations and propose strategies that could lead to wider acceptance of autosegmentation in routine clinical practice. The authors conclude that currently, autosegmentation technology in RT planning is an efficient tool for the clinicians to provide them with a good starting point for review and adjustment. Modern hardware platforms including GPUs allow most of the autosegmentation tasks to be done in a range of a few minutes. In the nearest future, improvements in CT-based autosegmentation tools will be achieved through standardization of imaging and contouring protocols. In the longer term, the authors expect a wider use of multimodality approaches and better understanding of correlation of imaging with biology and pathology.

Automatic segmentation of head and neck CT images for radiotherapy treatment planning using multiple atlases, statistical appearance models, and geodesic active contours.

PURPOSE: Accurate delineation of organs at risk (OARs) is a precondition for intensity modulated radiation therapy. However, manual delineation of OARs is time consuming and prone to high interobserver variability. Because of image artifacts and low image contrast between different structures, however, the number of available approaches for autosegmentation of structures in the head-neck area is still rather low. In this project, a new approach for automated segmentation of head-neck CT images that combine the robustness of multiatlas-based segmentation with the flexibility of geodesic active contours and the prior knowledge provided by statistical appearance models is presented. METHODS: The presented approach is using an atlas-based segmentation approach in combination with label fusion in order to initialize a segmentation pipeline that is based on using statistical appearance models and geodesic active contours. An anatomically correct approximation of the segmentation result provided by atlas-based segmentation acts as a starting point for an iterative refinement of this approximation. The final segmentation result is based on using model to image registration and geodesic active contours, which are mutually influencing each other. RESULTS: 18 CT images in combination with manually segmented labels of parotid glands and brainstem were used in a leave-one-out cross validation scheme in order to evaluate the presented approach. For this purpose, 50 different statistical appearance models have been created and used for segmentation. Dice coefficient (DC), mean absolute distance and max. Hausdorff distance between the autosegmentation results and expert segmentations were calculated. An average Dice coefficient of DC = 0.81 (right parotid gland), DC = 0.84 (left parotid gland), and DC = 0.86 (brainstem) could be achieved. CONCLUSIONS: The presented framework provides accurate segmentation results for three important structures in the head neck area. Compared to a segmentation approach based on using multiple atlases in combination with label fusion, the proposed hybrid approach provided more accurate results within a clinically acceptable amount of time.

Interactive medical image segmentation using PDE control of active contours.

Segmentation of injured or unusual anatomic structures in medical imagery is a problem that has continued to elude fully automated solutions. In this paper, the goal of easy-to-use and consistent interactive segmentation is transformed into a control synthesis problem. A nominal level set partial differential equation (PDE) is assumed to be given; this open-loop system achieves correct segmentation under ideal conditions, but does not agree with a human expert's ideal boundary for real image data. Perturbing the state and dynamics of a level set PDE via the accumulated user input and an observer-like system leads to desirable closed-loop behavior. The input structure is designed such that a user can stabilize the boundary in some desired state without needing to understand any mathematical parameters. Effectiveness of the technique is illustrated with applications to the challenging segmentations of a patellar tendon in magnetic resonance and a shattered femur in computed tomography.

Hip fracture discrimination from dual-energy X-ray absorptiometry by statistical model registration.

Although the areal Bone Mineral Density (BMD) measurements from dual-energy X-ray absorptiometry (DXA) are able to discriminate between hip fracture cases and controls, the femoral strength is largely determined by the 3D bone structure. In a previous work a statistical model was presented which parameterizes the 3D shape and BMD distribution of the proximal femur. In this study the parameter values resulting from the registration of the model onto DXA images are evaluated for their hip fracture discrimination ability with respect to regular DXA derived areal BMD measurements. The statistical model was constructed from a large database of QCT scans of females with an average age of 67.8 ± 17.0 years. This model was subsequently registered onto the DXA images of a fracture and control group. The fracture group consisted of 175 female patients with an average age of 66.4 ± 9.9 years who suffered a fracture on the contra lateral femur. The control group consisted of 175 female subjects with an average age of 65.3 ± 10.0 years and no fracture history. The discrimination ability of the resulting model parameter values, as well as the areal BMD measurements extracted from the DXA images were evaluated using a logistic regression analysis. The area under the receiver operating curve (AUC) of the combined model parameters and areal BMD values was 0.840 (95% CI 0.799-0.881), whilst using only the areal BMD values resulted in an AUC of 0.802 (95% CI 0.757-0.848). These results indicate that the discrimination ability of the areal BMD values is improved by supplementing them with the model parameter values, which give a more complete representation of the subject specific shape and internal bone distribution. Thus, the presented method potentially allows for an improved hip fracture risk estimation whilst maintaining DXA as the current standard modality.

3-D graph cut segmentation with Riemannian metrics to avoid the shrinking problem.

Though graph cut based segmentation is a widely-used technique, it is known that segmentation of a thin, elongated structure is challenging due to the "shrinking problem". On the other hand, many segmentation targets in medical image analysis have such thin structures. Therefore, the conventional graph cut method is not suitable to be applied to them. In this study, we developed a graph cut segmentation method with novel Riemannian metrics. The Riemannian metrics are determined from the given "initial contour," so that any level-set surface of the distance transformation of the contour has the same surface area in the Riemannian space. This will ensure that any shape similar to the initial contour will not be affected by the shrinking problem. The method was evaluated with clinical CT datasets and showed a fair result in segmenting vertebral bones.