This PDF file contains the front matter associated with SPIE Proceedings Volume 8669, including the Title Page, Copyright Information, Table of Contents, and the Conference Committee listing.

Markov random field (MRF) based approaches are extensively used in image segmentation applications, which often produce segmentation results with a boundary of minimal length/surface and tending to pass along image edges, yet affected by boundary shrinkage or bias in the absence of proper image edge information to drive the segmentation. In this paper, we propose a novel curvature re-weighted boundary smoothing term and introduce a new convex optimization-based contour/surface evolution method for medical image segmentation. The proposed curvature-based term generates the optimal solution with low curvatures and helps to avoid boundary shrinkage and bias. This is particularly useful for segmenting medical images, in which noisy and poor image quality exists widely and the shapes of anatomical objects are often smooth and even convex. Moreover, a new convex optimization-based contour evolution method is applied to propagate the initial contour to the object of interest efficiently and robustly. Distinct from the traditional methods for contour evolution, the proposed algorithm provides a fully time-implicit contour evolution scheme, which allows a large evolution step-size to significantly speed up convergence. It also propagates the contour to its globally optimal position during each discrete time-frame, which improves the algorithmic robustness to noise and poor initialization. The fast continuous max-flow-based algorithm for contour evolution is implemented on a commercially available graphics processing unit (GPU) to achieve a high computational performance. Experimental results for both synthetic and 2D/3D medical images showed that the proposed approach generated segmentation results efficiently and increased the accuracy and robustness of segmentation by avoiding segmentation shrinkage and bias.

Fluorescence in situ hybridization (FISH) dot counting is the process of enumerating chromosomal abnormalities in interphase cell nuclei. This process is widely used in many areas of biomedical research and diagnosis. We present a generic and fully automatic algorithm for cell-level counting of FISH dots in 2-D fluorescent images. Our proposed algorithm starts by segmenting cell nuclei in DAPI stained images using a 2-D wavelet based segmentation algorithm. Nuclei segmentation is followed by FISH dot detection and counting, which consists of three main steps. First, image pre-processing where median and top-hat filters are used to clean image noise, subtract background and enhance the contrast of the FISH dots. Second, FISH dot detection using a multi-level h-minima transform approach that accounts for the varying image contrast. Third, FISH dot counting where clustered FISH dots are separated using a local maxima detection-based method followed by FISH dot size filtering based on constraints to account for large connected components of tightly-clustered dots. To quantitatively assess the performance of our proposed FISH dot counting algorithm, automatic counting results were compared to manual counts of 880 cells selected from 19 invasive ductal breast carcinoma samples exhibiting varying degrees of Human Epidermal Growth Factor Receptor 2 (HER2) expression. Cell-level dot counting accuracy was assessed using two metrics: cell classification agreement and dot-counting match. Our automatic results gave an overall cell-by-cell classification agreement of 88% and an overall accuracy of 81%.

With the rapid development of microscopy for cell imaging, there is a strong and growing demand for image analysis software to quantitatively study cell morphology. Automatic cell segmentation is an important step in image analysis. Despite substantial progress, there is still a need to improve the accuracy, efficiency, and adaptability to different cell morphologies. In this paper, we propose a fully automatic method for segmenting cells in uorescence images of conuent cell monolayers. This method addresses several challenges through a combination of ideas. 1) It realizes a fully automatic segmentation process by first detecting the cell nuclei as initial seeds and then using a multi-object geometric deformable model (MGDM) for final segmentation. 2) To deal with different defects in the uorescence images, the cell junctions are enhanced by applying an orderstatistic filter and principal curvature based image operator. 3) The final segmentation using MGDM promotes robust and accurate segmentation results, and guarantees no overlaps and gaps between neighboring cells. The automatic segmentation results are compared with manually delineated cells, and the average Dice coefficient over all distinguishable cells is 0:88.

Segmentation of brain structures in magnetic resonance images is an important task in neuro image analysis. Several papers on this topic have shown the benefit of supervised classification based on local appearance features, often combined with atlas-based approaches. These methods require a representative annotated training set and therefore often do not perform well if the target image is acquired on a different scanner or with a different acquisition protocol than the training images. Assuming that the appearance of the brain is determined by the underlying brain tissue distribution and that brain tissue classification can be performed robustly for images obtained with different protocols, we propose to derive appearance features from brain-tissue density maps instead of directly from the MR images. We evaluated this approach on hippocampus segmentation in two sets of images acquired with substantially different imaging protocols and on different scanners. While a combination of conventional appearance features trained on data from a different scanner with multi-atlas segmentation performed poorly with an average Dice overlap of 0.698, the local appearance model based on the new acquisition-independent features significantly improved (0.783) over atlas-based segmentation alone (0.728).

Image segmentation for the demarcation of pulmonary nodules in CT images is intrinsically an arduous task. The difficulty can be summarized into two aspects. Firstly, lung tumor can be various in terms of physical densities in pulmonary regions, implying the different interpretation as a mixture of GGO and solid nodules. Hence, processing of lung CT images may generally encounter tissue inhomogeneous problem. The second factor that complicates the task of nodule demarcation is the irregular shapes that most nodules are directly connected to other structures sharing the similar density profile. In this paper, an image segmentation framework is proposed by unifying the techniques of statistical region merging and conditional random field (CRF) with graph cut optimization to address the difficult problem of GGO nodules quantification in CT images. Different from traditional segmentation methods that use pixel-based approach such as region growing and morphological constraints, we employ a hierarchical segmentation tree to alleviate the effect of inhomogeneous attenuation. In addition to building perceptual prominent regions, we perform inference in CRF model based on restricting the pool of segmented regions. Following that, an inference CRF model is carried out to detect and localize individual object instances in CT images. The proposed algorithm is evaluated with four sets of manual delineations on 77 lung CT images. Incorporating with the efficiency and accuracy of pulmonary nodules segmentation method proposed in this paper, a computer aided system is hence feasible to develop related clinical application.

This paper presents a novel pipeline for the registration of diffusion tensor images (DTI) with large pathological variations to normal controls based on the use of a novel feature map derived from white matter (WM) fiber tracts. The research presented aims towards an atlas based DTI analysis of subjects with considerable brain pathologies such as tumors or hydrocephalus. In this paper, we propose a novel feature map that is robust against variations in WM fiber tract integrity and use these feature maps to determine a landmark correspondence using a 3D point correspondence algorithm. This correspondence drives a deformation field computed using Gaussian radial basis functions(RBF). This field is employed as an initialization to a standard deformable registration method like demons. We present early preliminary results on the registration of a normal control dataset to a dataset with abnormally enlarged lateral ventricles affected by fatal demyelinating Krabbe disease. The results are analyzed based on a regional tensor matching criterion and a visual assessment of overlap of major WM fiber tracts. While further evaluation and improvements are necessary, the results presented in this paper highlight the potential of our method in handling registration of subjects with severe WM pathology.

Shape, scale, orientation and position, the physical features associated with white matter DTI tracts, can, either individually or in combination, be used to define feature spaces. Recent work by Mani et al.¹ describes a Riemannian framework in which these joint feature spaces are considered. In this paper, we use the tools and metrics defined within this mathematical framework to study morphological changes due to disease progression. We look at sections of the anterior corpus callosum, which describes a deep arc along the mid-sagittal plane, and show how multiple sclerosis and normal control populations have different joint shape-orientation signatures.

The thalamus is a sub-cortical gray matter structure that relays signals between the cerebral cortex and midbrain. It can be parcellated into the thalamic nuclei which project to different cortical regions. The ability to automatically parcellate the thalamic nuclei could lead to enhanced diagnosis or prognosis in patients with some brain disease. Previous works have used diffusion tensor images (DTI) to parcellate the thalamus, using either tensor similarity or cortical connectivity as information driving the parcellation. In this paper, we propose a method that uses the diffusion tensors in a different way than previous works to guide a multiple object geometric deformable model (MGDM) for parcellation. The primary eigenvector (PEV) is used to indicate the homogeneity of fiber orientations. To remove the ambiguity due to the fact that the PEV is an orientation, we map the PEV into a 5D space known as the Knutsson space. An edge map is then generated from the 5D vector to show divisions between regions of aligned PEV's. The generalized gradient vector flow (GGVF) calculated from the edge map drives the evolution of the boundary of each nucleus. Region based force, balloon force, and curvature force are also employed to refine the boundaries. Experiments have been carried out on five real subjects. Quantitative measures show that the automated parcellation agrees with the manual delineation of an expert under a published protocol.

Major white matter (WM) pathways in the brain can be reconstructed in vivo using tractography on diffusion tensor imaging (DTI) data. Performing tractography using the native DTI data is often considered to produce more faithful results than performing it using the spatially normalized DTI obtained using highly non-linear transformations. However, tractography in the normalized DTI is playing an increasingly important role in population analyses of the WM. In particular, the emerging tract specific analyses (TSA) can benefit from tractography in the normalized DTI for statistical parametric mapping in specific WM pathways. It is well known that the preservation of tensor orientations at the individual voxel level is enforced in tensor based registrations. However small reorientation errors at individual voxel level can accumulate and could potentially affect the tractography results adversely. To our knowledge, there has been no study investigating the effects of normalization on consistency of tractography that demands non-local preservation of tensor orientations which is not explicitly enforced in typical DTI spatial normalization routines. This study aims to evaluate and compare tract reconstructions obtained using normalized DTI against those obtained using native DTI. Although tractography results have been used to measure and influence the quality of spatial normalization, the presented study addresses a distinct question: whether non-linear spatial normalization preserves even long-range anatomical connections obtained using tractography for accurate reconstructions of pathways. Our results demonstrate that spatial normalization of DTI data does preserve tract reconstructions of major WM pathways and does not alter the variance (individual differences) of their macro and microstructural properties. This suggests one can extract quantitative and shape properties efficiently from the tractography data in the normalized DTI for performing population statistics on major WM pathways.

We present a novel method for correction of geometric distortions arising from susceptibility artefacts in echo-planar MRI images that combines fieldmap and an image registration based correction technique in a unified framework. The geometric distortions arising from these artefacts lead to inaccurate alignment of images from different MRI techniques and hinders their joint analysis. A novel phase unwrapping algorithm is presented that can efficiently compute the B0 field inhomogeneity map as well as the confidence associated with the estimated solution. This information is used to adaptively drive a subsequent image registration step to further refine the results in low-confidence areas. The effectiveness of the proposed unified algorithm in correcting for geometric distortions due to susceptibility artefacts is demonstrated on interventional MRI EPI images.

Brain network analysis is a promising tool in studies of the human brain’s functional organization and neuropsychiatric disorders. For neuroimaging data based brain network analysis, network nodes are typically defined as distinct greymatter regions delineated by anatomical brain atlases, random parcellation of the brain space, or image voxels, resulting in brain network nodes with different spatial scales. As precise functional organization of the brain remains unclear, it is challenging to determine a proper spatial scale in practice. The brain network nodes defined anatomically or randomly do not necessarily possess desired properties of ideal brain network nodes, i.e., functional homogeneity within each individual node, functional distinctiveness across different nodes, and functional consistence of the same node across different subjects. To obtain a definition of brain network nodes with the desired properties, a brain parcellation method based on functional information is proposed to achieve a brain parcellation consistent across subjects and highly in agreement with the functional organization of the brain. Particularly, spatially contiguous voxel-wise functional information of the brain fMRI data recursively aggregate according to inter-voxel/region functional affinity from voxel level to coarser scales, resulting in a brain parcellation with a multi-level hierarchy. A trade-off between functional homogeneity and distinctiveness is determined by identifying the hierarchy level with network measures highly consistent across subjects. The proposed method has been validated on resting-state fMRI datasets for functional network analysis, and the results demonstrate that brain networks constructed with 200~500 nodes could achieve the highest inter-subject consistence.

Shape deformations and volumetric changes in the hippocampus and amygdala have previously been noted in Major Depressive Disorder (MDD). Unfortunately, these analyses are limited because relative shape and pose (rigid+scale transformation) information of multiple objects in brain are generally disregarded. We hypothesize that this information might complement studies of limbic structural deformation in MDD. We focus on changes in temporal (e.g., superior, middle and inferior temporal gyrus) and limbic (e.g., hippocampus and amygdala) lobes. Here, we use a multi-object statistical pose and shape model to analyze imaging data from young people with and without a depressive disorder. Nineteen individuals with a depressive disorder (mean age: 17.85) and twenty six healthy controls (age: 18) were enrolled in the study. A segmented atlas in MNI space has been used to segment hippocampus, amygdala, parahippocampal gyri, putamen, and the superior, inferior and middle temporal gyri in both hemispheres of the brain. Points on the surface of each structure were extracted and warped to each subjects’ structural MRI. These corresponding surface points were used within the analysis, to extract the pose and shape features. Pose and shape differences were detected between the two groups, such that second principal mode of pose variation (p = 0.022), and first principal mode of shape variation (p = 0.049) were found to differ significantly between the two groups.

Statistical shape modeling is considered as a backbone of image analysis, since shapes capture distinguishable geometrical properties of depicted objects and spatial relationships among the objects. In the field of medical image analysis, a shape allows segmentation and registration of complex and/or poorly visible structures, where geometrical information may have a more crucial role than pure intensity information. In this paper, we present a novel statistical shape model based on landmark positions and spatial relationships among landmarks. A given training set of images is first annotated by a set of landmarks, which represents the shape of the object of interest. In contrast to active shape (ASM) and appearance models (AAM), where a shape is a single object characterized by a system of eigenvectors, we describe a shape as a combination of distances and angles between landmarks. Finding a suitable combination of distances and angles is achieved by optimizing the representativeness of the model (i.e. the distances and angles must describe the shape and its plasticity), and complexity of the model (i.e. the number of distances and angles must be acceptable for practical applications). To generate a model that satisfies these conditions, the landmarks are first separated into clusters, which are then optimally connected. The optimal connections between clusters are generated by using the assignment problem. The obtained model combined with the game-theoretic framework was applied to segment lung fields from chest radiographs. The usage of such simplified model results on average in a 0.05 mm deterioration of the segmentation performance in terms of the symmetric mean boundary distance, and in a 3.3-times acceleration of the computational time.

Identification of vertebral deformations in two dimensions (2D) is a challenging task due to the projective nature of radiographic images and natural anatomical variability of vertebrae. By generating detailed three-dimensional (3D) anatomical images, computed tomography (CT) enables accurate measurement of vertebral deformations. We present a novel approach to quantitative vertebral morphometry (QVM) based on parametric modeling of the vertebral body shape in 3D. A detailed 3D representation of the vertebral body shape is obtained by automatically aligning a parametric 3D model to vertebral bodies in CT images. The parameters of the 3D model describe clinically meaningful morphometric vertebral body features, and QVM in 3D is performed by comparing the parameters to their statistical values. By applying statistical classification analysis, thresholds and parameters that best discriminate between normal and fractured vertebral bodies are determined. The proposed QVM in 3D was applied to 454 normal and 228 fractured vertebral bodies, yielding classification sensitivity of 92:5% at 7:5% specificity, with corresponding accuracy of 92:5% and precision of 86:1%. The 3D shape parameters that provided the best separation between normal and fractured vertebral bodies were the vertebral body height, and the inclination and concavity of both vertebral endplates. The described QVM in 3D is able to efficiently discriminate between normal and fractured vertebral bodies, and identify morphological cases (wedge, (bi)concavity, crush) and grades (1, 2, 3) of vertebral body deformations. It may be therefore valuable for diagnosing and predicting vertebral fractures in patients who are at risk of osteoporosis.

In this work we automatically segment the left ventricle (LV) in cardiac MR images in the end-diastole (ED) and end-systole (ES) phases using a novel approach that combines statistical and deterministic deformable models. A 3D Active Appearance Model (AAM) is used to segment the ED phase. The AAM texture model is trained on radial samples from gradient magnitude images to make the fitting process faster and more discriminative. A trained ED-to-ES shape correspondence model is used to map a given ED shape to an ES shape. Once the AAM model converges to a shape in ED, the correspondence model is used to get an approximate ES shape. We segment the LV in the ES phase by first fitting a deformable superquadric to the AAM converged shape (in ED) using data range forces and then tracking the LV using image and data range forces (for the ES shape obtained from correspondence model). We test our approach by performing leave-one-out training on a 35 patient datasets. The data comprises 19 normal patients and 16 patients having heart abnormalities (cardiomyopathy and myocardial infarction). The composition makes it a challenging data collection with significant shape variation. The performance of our method is evaluated by measuring the mismatch between automatically segmented and expert delineated contours using the Mean Perpendicular Distance (MPD) and Dice metrics. The average MPD is 2.6mm for ED and 3.7mm for ES (error mostly towards the apex and base). The average Dice is 0.9 for ED and 0.8 for ES. These results show good potential for clinical use.

This work addresses the challenging problem of parsing 2D radiographs into salient anatomical regions such as the left and right lungs and the heart. We propose the integration of an automatic detection of a constellation of landmarks via rejection cascade classifiers and a learned geometric constellation subset detector model with a multi-object active appearance model (MO-AAM) initialized by the detected landmark constellation subset. Our main contribution is twofold. First, we propose a recovery method for false positive and negative landmarks which allows to handle extreme ranges of anatomical and pathological variability. Specifically we (1) recover false negative (missing) landmarks through the consensus of inferences from subsets of the detected landmarks, and (2) choose one from multiple false positives for the same landmark by learning Gaussian distributions for the relative location of each landmark. Second, we train a MO-AAM using the true landmarks for the detectors and during test, initialize the model using the detected landmarks. Our model fitting allows simultaneous localization of multiple regions by encoding the shape and appearance information of multiple objects in a single model. The integration of landmark detection method and MO-AAM reduces mean distance error of the detected landmarks from 20.0mm to 12.6mm. We assess our method using a database of scout CT scans from 80 subjects with widely varying pathology.

Multiple sclerosis (MS) is a disease with heterogeneous evolution among the patients. Some classifications have been carried out according to either the clinical course or the immunopathological profiles. Epidemiological data and imaging are showing that MS is a two-phase neurodegenerative inflammatory disease. At the early stage it is dominated by focal inflammation of the white matter (WM), and at a later stage it is dominated by diffuse lesions of the grey matter and spinal cord. A Clinically Isolated Syndrome (CIS) is a first neurological episode caused by inflammation/demyelination in the central nervous system which may lead to MS. Few studies have been carried out so far about this initial stage. Better understanding of the disease at its onset will lead to a better discovery of pathogenic mechanisms, allowing suitable therapies at an early stage. We propose a new data processing framework able to provide an early characterization of CIS patients according to lesion patterns, and more specifically according to the nature of the inflammatory patterns of these lesions. The method is based on a two layers classification. Initially, the spatio-temporal lesion patterns are classified using a tensor-like representation. The discovered lesion patterns are then used to identify group of patients and their correlation to 15 months follow-up total lesion loads (TLL), which is so far the only image-based figure that can potentially infer future evolution of the pathology. We expect that the proposed framework can infer new prospective figures from the earliest imaging sign of MS since it can provide a classification of different types of lesion across patients.

Increasing attention has been focused on the estimation of the deformation of the endocardium to aid the diagnosis of cardiac malfunction. Landmark tracking can provide sparse, anatomically relevant constraints to help establish correspondences between images being tracked or registered. However, landmarks on the endocardium are often characterized by ambiguous appearance in cardiac MR images which makes the extraction and tracking of these landmarks problematic. In this paper we propose an automatic framework to select and track a sparse set of distinctive landmarks in the presence of relatively large deformations in order to capture the endocardial motion in cardiac MR sequences. To achieve this a sparse set of the landmarks is identified using an entropy-based approach. In particular we use singular value decomposition (SVD) to reduce the search space and localize the landmarks with relatively large deformation across the cardiac cycle. The tracking of the sparse set of landmarks is performed simultaneously by optimizing a two-stage Markov Random Field (MRF) model. The tracking result is further used to initialize registration based dense motion tracking. We have applied this framework to extract a set of landmarks at the endocardial border of the left ventricle in MR image sequences from 51 subjects. Although the left ventricle undergoes a number of different deformations, we show how the radial, longitudinal motion and twisting of the endocardial surface can be captured by the proposed approach. Our experiments demonstrate that motion tracking using sparse landmarks can outperform conventional motion tracking by a substantial amount, with improvements in terms of tracking accuracy of 20:8% and 19:4% respectively.

We present a method that utilizes registration displacement fields to perform accurate classification of magnetic resonance images (MRI) of the brain acquired from healthy individuals and patients diagnosed with multiple sclerosis (MS). Contrary to standard approaches, each voxel in the displacement field is treated as an independent feature that is classified individually. Results show that when used with a simple linear discriminant and majority voting, the approach is superior to using the displacement field with a single classifier, even when compared against more sophisticated classification methods such as adaptive boosting, random forests, and support vector machines. Leave-one-out cross-validation was used to evaluate this method for classifying images by disease, MS subtype (Acc: 77%-88%), and age (Acc: 96%-100%).

Dynamic contrast-enhanced MRI is a dynamic imaging technique that is now widely used for cancer imaging. Changes in tumour microvasculature are typically quantified by pharmacokinetic modelling of the contrast uptake curves. Reliable pharmacokinetic parameter estimation depends on the measurement of the arterial input function, which can be obtained from arterial blood sampling, or extracted from the image data directly. However, arterial blood sampling poses additional risks to the patient, and extracting the input function from MR intensities is not reliable. In this work, we propose to compute a perfusion CT based arterial input function, which is then employed for dynamic contrast enhanced MRI pharmacokinetic parameter estimation. Here, parameter estimation is performed simultaneously with intra-sequence motion correction by using nonlinear image registration. K_trans maps obtained with this approach were compared with those obtained using a population averaged arterial input function, i.e. Orton. The dataset comprised 5 rectal cancer patients, who had been imaged with both perfusion CT and dynamic contrast enhanced MRI, before and after the administration of a radiosensitising drug. K_trans distributions pre and post therapy were computed using both the perfusion CT and the Orton arterial input function. Perfusion CT derived arterial input functions can be used for pharmacokinetic modelling of dynamic contrast enhanced MRI data, when perfusion CT images of the same patients are available. Compared to the Orton model, perfusion CT functions have the potential to give a more accurate separation between responders and non-responders.

We present a novel algorithm for the registration of multiple temporally related point sets. Although our algorithm is derived in a general setting, our primary motivating application is coronary tree matching in multi-phase cardiac spiral CT. Our algorithm builds upon the fast, outlier-resistant Coherent Point Drift (CPD) algorithm, but incorporates temporal consistency constraints between the point sets, resulting in spatiotemporally smooth displacement fields. We preserve the speed and robustness of the CPD algorithm by using the technique of separable surrogates within an EM (Expectation-Maximization) optimization framework, while still minimizing a global registration cost function employing both spatial and temporal regularization. We demonstrate the superiority of our novel temporally consistent group-wise CPD algorithm over a straightforward pair-wise approach employing the original CPD algorithm, using coronary trees derived from both simulated and real cardiac CT data. In all the tested configurations and datasets, our method presents lower average error between tree landmarks compared to the pairwise method. In the worst case, the difference is around few micrometers but in the better case, our method divides by two the error from the pairwise method. This improvement is especially important for a dataset with numerous outliers. With a fixed set of parameter that has been tuned automatically, our algorithm yields better results than the original CPD algorithm which shows the capacity to register without a priori information on an unknown dataset.

X-ray exposure during image guided interventions can be important for the patient as well as for the medical staff. Therefore dose reduction is a major concern. Nevertheless, decreasing the dose per image affects significantly the image quality. As a matter of fact, this tends to increase the noise and reduce the contrast. Hence, we propose a new and efficient method to reduce the noise in low dose fluoroscopic sequences. Many studies in that domain have been proposed implementing either multi-scale approaches using wavelet with its derivatives or using filters in the direct space. Our work is based on a spatio-temporal denoising filter using the curvelet transform. Indeed, this sparse transform represents well smooth images with edges and can be applied to fluoroscopic images in order to achieve robust denoising performances. Therefore, we propose to combine a temporal recursive filter with a spatial curvelet filter. Our work is focused on the use of the statistical dependencies between the curvelet coefficients in order to optimize the threshold function. Determining the correlation among coefficients allows to detect which coefficients represent the relevant signal. Thus, our method allows to diminish or even to erase curvelet-like artefacts. The performances and robustness of the proposed method are assessed both on synthetic and real low dose sequences (ie: 20 nGy/frame).

Glaucoma is one of the major causes of blindness worldwide. One important structural parameter for the diagnosis and management of glaucoma is the cup-to-disc ratio (CDR), which tends to become larger as glaucoma progresses. While approaches exist for segmenting the optic disc and cup within fundus photographs, and more recently, within spectral-domain optical coherence tomography (SD-OCT) volumes, no approaches have been reported for the simultaneous segmentation of these structures within both modalities combined. In this work, a multimodal pixel-classification approach for the segmentation of the optic disc and cup within fundus photographs and SD-OCT volumes is presented. In particular, after segmentation of other important structures (such as the retinal layers and retinal blood vessels) and fundus-to-SD-OCT image registration, features are extracted from both modalities and a k-nearest-neighbor classification approach is used to classify each pixel as cup, rim, or background. The approach is evaluated on 70 multimodal image pairs from 35 subjects in a leave-10%-out fashion (by subject). A significant improvement in classification accuracy is obtained using the multimodal approach over that obtained from the corresponding unimodal approach (97.8% versus 95.2%; p < 0:05; paired t-test).

Ultrasound (US) is a versatile, low cost, real-time, widely available imaging modality. Manual segmentation for volumetric US measurements can be difficult and very time consuming, requiring slice-by-slice segmentations. However, automatic segmentation of ultrasound images can prove challenging due to the presence of speckle, attenuation, missing boundaries, signal dropouts, and artefacts. Semi-automatic segmentation techniques can improve the speed and accuracy of such measurements, taking advantage of clinical expertise while allowing user interaction. This paper presents a novel solution for interactive image segmentation on B-mode ultrasound images. The proposed method builds on the Live Wire framework and introduces two new sets of Live Wire costs, namely a Feature Asymmetry (FA) cost to localise edges and a weak shape constraint cost to aid the selection of appropriate boundaries in the presence of missing information or artefacts. The resulting semi-automatic segmentation method follows edges based on structural relevance rather than intensity gradients, adapting the method to ultrasound images, where the object boundaries are normally fuzzy. The new method is applied in the context of fetal arm adipose tissue quantification, the adipose tissue being an indicator of the fetal nutritional state. A quantitative and qualitative evaluation is performed with respect to related segmentation techniques. The method was tested on 48 manually segmented ultrasound images of the fetal arm across gestation, showing similar accuracy to the intensity-based Live Wire approach but superior repeatability while requiring significantly less time and user interaction.

An automatic framework is proposed to segment right ventricle on ultrasound images. This method can automatically segment both epicardial and endocardial boundaries from a continuous echocardiography series by combining sparse matrix transform (SMT), a training model, and a localized region based level set. First, the sparse matrix transform extracts main motion regions of myocardium as eigenimages by analyzing statistical information of these images. Second, a training model of right ventricle is registered to the extracted eigenimages in order to automatically detect the main location of the right ventricle and the corresponding transform relationship between the training model and the SMT-extracted results in the series. Third, the training model is then adjusted as an adapted initialization for the segmentation of each image in the series. Finally, based on the adapted initializations, a localized region based level set algorithm is applied to segment both epicardial and endocardial boundaries of the right ventricle from the whole series. Experimental results from real subject data validated the performance of the proposed framework in segmenting right ventricle from echocardiography. The mean Dice scores for both epicardial and endocardial boundaries are 89.1%±2.3% and 83.6±7.3%, respectively. The automatic segmentation method based on sparse matrix transform and level set can provide a useful tool for quantitative cardiac imaging.

Optical coherence tomography (OCT) has become one of the most common tools for diagnosis of retinal abnormalities. Both retinal morphology and layer thickness can provide important information to aid in the differential diagnosis of these abnormalities. Automatic segmentation methods are essential to providing these thickness measurements since the manual delineation of each layer is cumbersome given the sheer amount of data within each OCT scan. In this work, we propose a new method for retinal layer segmentation using a random forest classifier. A total of seven features are extracted from the OCT data and used to simultaneously classify nine layer boundaries. Taking advantage of the probabilistic nature of random forests, probability maps for each boundary are extracted and used to help refine the classification. We are able to accurately segment eight retinal layers with an average Dice coefficient of 0:79±0:13 and a mean absolute error of 1:21±1:45 pixels for the layer boundaries.

Carotid atherosclerosis is the major cause of ischemic stroke, a leading cause of mortality and disability. Many research studies have been carried out on how to quantitatively evaluate local arterial effects of potential carotid disease treatments. In this paper, the atorvastatin effect evaluation on atherosclerosis plaques are classified based on various shape and texture features extracted from ultrasound images. First, images of atherosclerotic lesions were extracted manually from ultrasound images by an expert physician. After analysis, 26 shape and 85 texture characteristics, and vessel wall volume (VWV) percent of change, were extracted and calculated from atherosclerotic lesions. Among these, to make the method convenient and exact enough, effective features and VWV percent of change, were selected for drug treatment effect evaluation by physician. Finally, a support vector machine (SVM) classifier was utilized to classify atherosclerosis plaques between atorvastatin group and placebo group. The leave-one-case-out protocol was utilized on a database of 768 carotid ultrasound images of 12 patients (5 subjects of placebo group and 7 subjects of atorvastatin group) for evaluation. The classification results showed overall accuracy 91.67%, sensitivity 95.56%, specificity 86.16%; positive predictive value 90.72%, negative predictive value 93.20%, Matthew’s correlation coefficient 82.81%, Youden’s index 81.72%. And the receiver operating characteristic (ROC) curve in our test also performed well. The experimental results also demonstrate that classification using the combined features has higher accuracy than that only using shape/texture feature or VWV percent of change. The proposed method can be used for the statins effect evaluation, especially when patients are treated with drugs, and further be developed as a beneficial tool for facilitating a physician’s diagnosis of the atherosclerosis.

Motion, like tumor movement due to respiration, constitutes a major problem in radiotherapy and/or diagnostics. A common idea to compensate for the motion in 4D imaging, is to invoke a registration strategy, which aligns the images over time. This approach is especially challenging if real time processing of the data and robustness with respect to noise and acquisition errors is required. To this end, we present a novel method which is based only on selected image features and uses a probabilistic approach to compute the wanted transformations of the 3D images. Moreover, we restrict the search space to rotation, translation and scaling. In an initial phase, landmarks in the first image of the series have to be identified, which are in the course of the scheme automatically transferred to the next image. To find the associated transformation parameters, a probabilistic approach, based on factored sampling, is invoked. We start from a state set containing a fixed number of different candidate parameters whose probabilities are approximated based on the image information at the landmark positions. Subsequent time frames are analyzed by factored sampling from this state set and by superimposing a stochastic diffusion term on the parameters. The algorithm is successfully applied to clinical 4D CT data. Landmarks have been placed manually to mark the tumor or a similar structure in the initial image whose position is then tracked over time. We achieve a processing rate of up to 12 image volumes per second. The accuracy of the tracking after five time steps is measured based on expert placed landmarks. We achieve a mean landmark error of less than 2 mm in each dimension in a region with radius of 25 mm around the target structure.

Treatment of tumor sites affected by respiratory motion requires knowledge of the position and the shape of the tumor and the surrounding organs during breathing. As not all structures of interest can be observed in real-time, their position needs to be predicted from partial information (so-called surrogates) like motion of diaphragm, internal markers or patients surface. Here, we present an approach to model respiratory lung motion and predict the position and shape of the lungs from surrogates. 4D-MRI lung data of 10 healthy subjects was acquired and used to create a model based on Principal Component Analysis (PCA). The mean RMS motion ranged from 1.88 mm to 9.66 mm. Prediction was done using a Bayesian approach and an average RMSE of 1.44 mm was achieved.

Pulmonary fissures are important landmarks for automated recognition of lung anatomy and need to be detected as a pre-processing step. We propose a derivative of stick (DoS) filter for pulmonary fissures detection in thoracic CT scans by considering their thin curvilinear shape across multiple transverse planes. Based on a stick decomposition of a local rectangular neighborhood, a nonlinear derivative operator perpendicular to each stick is defined. Then, combining with a standard deviation of the intensity along the stick, the composed likelihood function will take a strong response to fissure-like bright lines, and tends to suppress undesired structures including large vessels, step edges and blobs. Applying the 2D filter sequentially to the sagittal, coronal and axial slices, an approximate 3D co-planar constraint is implicitly exerted through the cascaded pipeline, which helps to further eliminate non-fissure tissues. To generate a clear fissure segmentation, we adopt a connected component based post-processing scheme, combined with a branch-point finding algorithm to disconnect the residual adjacent clutters from the fissures. The performance of our filter has been verified in experiments with a 23 patients dataset, where pathologies to different extents are included. The DoS filter compared favorably with prior algorithms.

Four-dimensional CT scans provides valuable motion information of patient throughout different respiratory phases. PET, on the other hand, provides functional information about tumor, which differentiate tumor from normal tissue effectively. However, manually contouring structures of interest on 4D CT is prohibitively tedious due to the large amount of data. In this paper, we propose an automatic method to segment lung tumor simultaneously for 4D CT scans in all phases and PET scan. The problem is modeled as an optimization problem based on Markov Random Fields (MRF) which involves region, boundary terms and a regularization term between PET and CT scans. The problem is solved optimally by computing a single max flow in a properly constructed graph. As far as the authors know, this is the first work in simultaneously segmenting tumor in 4D CT while incorporating PET information. Experiments on 3 lung cancer patients are conducted. The average Dice coefficient is improved from 0.680 to 0.791 compared to segmenting tumor volume in 4D CT phase by phase without incorporating PET information. The proposed method is efficient in terms of running time since the method only requires computing a max flow for which efficient algorithm exists. The memory consumption is linearly scalable with respect to number of 4D CT phases, which enables our method to handle multiple 4D CT phases with reasonable memory consumption.

Lung lobe segmentation is clinically important for disease classification, treatment and follow-up of pulmonary diseases. Diseases such as tuberculosis and silicolis typically present in specific lobes i.e. almost exclusively the upper ones. However, the fissures separating different lobes are often difficult to detect because of their variable shape, appearance and low contrast in computed tomography images. In addition, a substantial fraction of patients have missing or incomplete fissures. To solve this problem, several methods have been employed to interpolate incomplete or missed fissures. For example, Pu et al. used an implicit surface fitting with different radial basis functions; Ukil et al. apply fast marching methods; and Ross et al. used an interactive thin plate spline (TPS) interpolation where the user selects the points that will be used to compute the fissure interpolation via TPS. In our study, results of an automated fissure detection method based on a plate-filter as well points derived from vessels were fed into an a robust TPS interpolation that ultimately defined the lobes. To improve the selection of detected points, we statistically determined the areas where fissures are localized from 19 data-sets. These areas were also used to constrain TPS fitting so it reflected the expected shape and orientation of the fissures, hence improving result accuracy. Regions where the detection step provided low response were replaced by points derived from a distance-to-vessels map. The error, defined as the Euclidian mean distance between ground truth points and the TPS fitted fissures, was computed for each dataset to validate our results. Ground truth points were defined for both exact fissure locations and approximate fissure locations (when the fissures were not clearly visible). The mean error was 5.64±4.83 mm for the exact ground truth points, and 10.01 ± 8.23 mm for the approximate ground truth points.

Lung registration in thoracic CT scans has received much attention in the medical imaging community. Possible applications range from follow-up analysis, motion correction for radiation therapy, monitoring of air flow and pulmonary function to lung elasticity analysis. In a clinical environment, runtime is always a critical issue, ruling out quite a few excellent registration approaches. In this paper, a highly efficient variational lung registration method based on minimizing the normalized gradient fields distance measure with curvature regularization is presented. The method ensures diffeomorphic deformations by an additional volume regularization. Supplemental user knowledge, like a segmentation of the lungs, may be incorporated as well. The accuracy of our method was evaluated on 40 test cases from clinical routine. In the EMPIRE10 lung registration challenge, our scheme ranks third, with respect to various validation criteria, out of 28 algorithms with an average landmark distance of 0.72 mm. The average runtime is about 1:50 min on a standard PC, making it by far the fastest approach of the top-ranking algorithms. Additionally, the ten publicly available DIR-Lab inhale-exhale scan pairs were registered to subvoxel accuracy at computation times of only 20 seconds. Our method thus combines very attractive runtimes with state-of-the-art accuracy in a unique way.

4D imaging becomes increasingly important in clinical practice. Its use in diagnostics and therapy planning usually requires the application of non-linear registration techniques. The reliability of information derived from the computed transformations is directly dependent on the registration accuracy. Ideally, this accuracy should be evaluated on a patient- and data-specific level { which requires appropriate evaluation criteria and procedures. A standard approach for evaluation of non-linear registration accuracy is to compute a landmark- or point-based registration error by means of manually detected landmark correspondences in the images to register, with the landmarks being anatomically characteristic points. Manual detection of such points is, however, time-consuming and error-prone. In this contribution, different operators for automatic landmark detection and a block matching strategy for landmark propagation in 4D image sequences (here: 4D lung CT, 4D liver MRT) are proposed and evaluated. It turns out that the so-called Förstner-Rohr operators perform best for detection of anatomically characteristic points and that the proposed propagation strategy ensures a robust transfer of these landmarks between the images. The automatically detected landmark correspondences are then used to evaluate the accuracy of different registration approaches (in total 48 variants) applied for registering 4D lung CT data. The resulting registration error values are compared to errors obtained by manually detected landmark pairs. It is shown that derived statements concerning differences in accuracy of the registration approaches are identical for both the manually and the automatically detected landmark sets.

Some of the hardest problems in deformable image registration are problems where large anatomical differences occur between image acquisitions (e.g. large deformations due to images acquired in prone and supine positions and (dis)appearing structures between image acquisitions due to surgery). In this work we developed and studied, within a previously introduced multi-objective optimization framework, a dual-dynamic transformation model to be able to tackle such hard problems. This model consists of two non-fixed grids: one for the source image and one for the target image. By not requiring a fixed, i.e. pre-determined, association of the grid with the source image, we can accommodate for both large deformations and (dis)appearing structures. To find the transformation that aligns the source with the target image we used an advanced, powerful model-based evolutionary algorithm that exploits features of a problem’s structure in a principled manner via probabilistic modeling. The actual transformation is given by the association of coordinates with each point in the two grids. Linear interpolation inside a simplex was used to extend the correspondence (i.e. transformation) as found for the grid to the rest of the volume. As a proof of concept we performed tests on both artificial and real data with disappearing structures. Furthermore, the case of prone-supine image registration for 2D axial slices of breast MRI scans was evaluated. Results demonstrate strong potential of the proposed approach to account for large deformations and (dis)appearing structures in deformable image registration.

In this paper we show how to use the approximate bilateral symmetry of the brain with respect to its interhemispheric fissure for intra-subject (rigid-body) mono- and multimodal 3D image registration. We propose to define and compute an approximate symmetry plane in the two images to register and to use these two planes as constraints in the registration problem. This 6-parameter problem is thus turned into three successive 3-parameter problems. Our hope is that the lower dimension of the parameter space makes these three subproblems easier and faster to solve than the initial one. We implement two algorithms to solve these three subproblems in the exact same way, within a common intensity-based framework using mutual information as the similarity measure. We compare this symmetry-based strategy with the standard approach (i.e. direct estimation of a 6-parameter rigid-body transformation), also implemented within the same framework, using synthetic and real datasets. We show our symmetry-based method to achieve subvoxel accuracy with better robustness and larger capture range than the standard approach, while being slightly less accurate and slower. Our method also succeeds in registering clinical MR and PET images with a much better accuracy than the standard approach. Finally, we propose a third strategy to decrease the run time of the symmetry-based approach and we give some ideas, to be tested in future works, on how to improve its accuracy.

CT colonography interpretation is difficult and time-consuming because fecal residue or fluid can mimic or obscure polyps, leading to diagnostic errors. To compensate for this, it is normal practice to obtain CT data with the patient in prone and supine positions. Repositioning redistributes fecal residue and colonic gas; fecal residue tends to move, while fixed mural pathology does not. The cornerstone of competent interpretation is the matching of corresponding endoluminal locations between prone and supine acquisitions. Robust and accurate automated registration between acquisitions should lead to faster and more accurate detection of colorectal cancer and polyps. Any directional bias when registering the colonic surfaces could lead to incorrect anatomical correspondence resulting in reader error. We aim to reduce directional bias and so increase robustness by adapting a cylindrical registration algorithm to penalize inverse-consistency error, using a symmetric optimization. Using 17 validation cases, the mean inverse-consistency error was reduced significantly by 86%, from 3.3 mm to 0.45 mm. Furthermore, we show improved alignment of the prone and supine colonic surfaces, evidenced by a reduction in the mean-of-squared-differences by 43% overall. Mean registration error, measured at a sparse set of manually selected reference points, remained at the same level as the non-symmetric method (no significant differences). Our results suggest that the inverse-consistent symmetric algorithm performs more robustly than non-symmetric implementation of B-spline registration.

Statistical imaging atlases allow for integration of information from multiple patient studies collected across different image scales and modalities, such as multi-parametric (MP) MRI and histology, providing population statistics regarding a specific pathology within a single canonical representation. Such atlases are particularly valuable in the identification and validation of meaningful imaging signatures for disease characterization in vivo within a population. Despite the high incidence of prostate cancer, an imaging atlas focused on different anatomic structures of the prostate, i.e. an anatomic atlas, has yet to be constructed. In this work we introduce a novel framework for MRI atlas construction that uses an iterative, anatomically constrained registration (AnCoR) scheme to enable the proper alignment of the prostate (Pr) and central gland (CG) boundaries. Our current implementation uses endorectal, 1.5T or 3T, T2-weighted MRI from 51 patients with biopsy confirmed cancer; however, the prostate atlas is seamlessly extensible to include additional MRI parameters. In our cohort, radical prostatectomy is performed following MP-MR image acquisition; thus ground truth annotations for prostate cancer are available from the histological specimens. Once mapped onto MP-MRI through elastic registration of histological slices to corresponding T2-w MRI slices, the annotations are utilized by the AnCoR framework to characterize the 3D statistical distribution of cancer per anatomic structure. Such distributions are useful for guiding biopsies toward regions of higher cancer likelihood and understanding imaging profiles for disease extent in vivo. We evaluate our approach via the Dice similarity coefficient (DSC) for different anatomic structures (delineated by expert radiologists): Pr, CG and peripheral zone (PZ). The AnCoR-based atlas had a CG DSC of 90.36%, and Pr DSC of 89.37%. Moreover, we evaluated the deviation of anatomic landmarks, the urethra and veromontanum, and found 3.64 mm and respectively 4.31 mm. Alternative strategies that use only the T2-w MRI or the prostate surface to drive the registration were implemented as comparative approaches. The AnCoR framework outperformed the alternative strategies by providing the lowest landmark deviations.

Mouse lung models facilitate the study of the pathogenesis of various pulmonary diseases such as infections and inflammatory diseases. The co-registration of ex vivo histological data and pre-excised magnetic resonance imaging (MRI) in preclinical mouse models would allow for determination and validation of imaging signatures for different pathobiologies within the lung. While slice-based co-registration could be used, this approach assumes that (a) slice correspondences between the two different modalities exist, and (b) finding slice correspondences often requires the intervention of an expert and is time consuming. A more practical approach is to first reconstruct the 3D histological volume from individual slices, then perform 3D registration with the MR volume. Before the histological reconstruction, image registration is required to compensate for geometric differences between slices. Pairwise algorithms work by registering pairs of successive slices. However, even if successive slices are registered reasonably well, the propagation of registration errors over slices can yield a distorted volumetric reconstruction significantly different in shape from the shape of the true specimen. Groupwise registration can reduce the error propagation by considering more than two successive images during the registration, but existing algorithms are computationally expensive. In this paper, we present an efficient groupwise registration approach, which yields consistent volumetric reconstruction and yet runs equally fast as pairwise registration. The improvements are based on 1) natural gradient which speeds up the transform warping procedure and 2) efficient optimization of the cost function of our groupwise registration. The strength of the natural gradient technique is that it could help mitigate the impact of the uncertainties of the gradient direction across multiple template slices. Experiments on two mouse lung datasets show that compared to pairwise registration, our groupwise approach runs faster in terms of registration convergence, and yields globally more consistent reconstruction.

Respiratory motion is a major source of error in radiation treatment of thoracic and abdominal tumors. State-of-the-art motion-adaptive radiation therapy techniques are usually guided by external breathing signals acting as surrogates for the internal motion of organs and tumors. Assuming a relationship between the surrogate measurements and the internal motion patterns, which are usually described by non-linear transformations, correspondence models can be defined and used for surrogate-based motion estimation. In this contribution, a diffeomorphic motion estimation framework based on standard multivariate linear regression is extended by subspace-based approaches like principal component analysis, partial least squares, and canonical correlation analysis. These methods aim at exploiting the hidden structure of the training data to improve the use of the information provided by high-dimensional surrogate and internal motion representations. A quantitative evaluation carried out on 4D CT data sets of 10 lung tumor patients shows that subspace-based approaches are able to significantly improve the mean estimation accuracy when compared to standard multivariate linear regression.

As medical imaging moves from qualitative assessment to quantitative analysis based on bio-markers, calibration of imaging data becomes critically important. In computed tomography (CT) images values are scaled to Hounsfield units (HU), despite this, the measurements can vary due to differences in machine-level calibration and patient size. One way to ensure proper calibration at the image level is to include a phantom containing objects of known HU values in each scan so that each image can be calibrated individually by the values measured from the phantom regions within the image. This introduces a need to extract phantom measurements from each image. Given a reasonable starting point manually or heuristically, this is a straightforward segmentation problem because the phantom regions are well defined and the values are relatively uniform. However, the problem becomes challenging if the requirement is a fully automated method that is robust across variations of phantoms and that can exclude images without phantoms. In this paper, we describe a probabilistic model-based approach to tackling this problem. We use the constellation model framework first proposed by Burt el al. to represent a phantom as composed of a number of parts and determine the existence and localization of the phantom in a probabilistic sense based on the detection of candidate parts. This model based approach allows us to formally describe variations in phantom design and handle missing parts caused by phantom regions similar to the background. Initial results on 100 CT studies from a longitudinal cardiovascular study are encouraging.

The detection of anatomical landmarks (LMs) often plays a key role in medical image analysis. In our previous study, we reported an automatic LM detection method for CT images. Despite its high detection sensitivity, the distance errors of the detection results for some LMs were relatively large as they sometimes exceeded 10 mm. Naturally, it is desirable to minimize LM detection error, especially when the LM detection results are used in image analysis tasks such as image segmentation. In this study, we introduce a novel method of coarse-to-fine localization to increase accuracy, which refines the LM positions detected by our previous method. The proposed LM localization is performed by both multiscale local image pattern recognition and likelihood estimation from prior knowledge of the spatial distribution of multiple LMs. Classifier ensembles for recognizing local image patterns are trained by the cost-sensitive MadaBoost. The cost of each sample is altered depending on its distance from the ground truth LM position. The spatial LM distribution likelihood, calculated from a statistical model of inter-landmark distances between all LM pairs, is also used in the localization. The evaluation experiment was performed with 15 LMs in 39 CT images. The average distance error of the pre-detected LM position was improved by 2.05 mm by the proposed localization method. The proposed method was shown to be effective for reducing LM detection error.

Labeling of cerebral vasculature is important for characterization of anatomical variation, quantification of brain morphology with respect to specific vessels, and inter-subject comparisons of vessel properties and abnormalities. We propose an automated method to label the anterior portion of cerebral arteries using a statistical inference method on the Bayesian network representation of the vessel tree. Our approach combines the likelihoods obtained from a random forest classifier trained using vessel centerline features with a belief propagation method integrating the connection probabilities of the cerebral artery network. We evaluate our method on 30 subjects using a leave-one-out validation, and show that it achieves an average correct vessel labeling rate of over 92%.

In this study, a pattern recognition-based framework is presented to automatically segment the complete cerebral vasculature from 4D Computed Tomography (CT) patient data. Ten consecutive patients whom were admitted to our hospital on a suspicion of ischemic stroke were included in this study. A background mask and bone mask were calculated based on intensity thresholding and morphological operations, and the following six image features were proposed: 1) a subtraction image of a subtraction image consisting of timing-invariant CTA and non-constrast CT, 2) the area under the curve of a gamma variate function fitted to the tissue curves, 3-5) three optimized parameter values of this gamma variate function, and 6) a vessel likeliness function. After masking bone and background, these features were used to train a linear discriminant voxel classifier (LDC) on regions of interest (ROIs), which were annotated in soft tissue (white matter and gray matter) and vessels by an expert observer. The LDC was trained in a leave-one-out manner in which 9 patients tissue ROIs were used for training and the remaining patient tissue ROIs were used for testing the classifier. To evaluate the frame work, for each training cycle the accuracy was calculated by dividing the true positives and negatives by the true positives and negatives and false positives and negatives. The resulting averaged accuracy was 0:985±0:014 with a range of 0:957 to 0:999.

Modern volumetric imaging techniques such as CT or MRI, aid in the understanding of a patient's anatomy and pathologies. Depending on the medical use case, various anatomical structures are of interest. Blood vessels play an important role in several applications, e.g. surgical planning. Manual delineation of blood vessels in volumetric images is error prone and time consuming. Automated vessel segmentation is a challenging problem due to acquisition-dependent problems such as noise, contrast, spatial resolution, and artifacts. In this paper, a vessel segmentation method is presented that combines a wavefront propagation technique with Hessian-based vessel enhancement. The latter has proven its usefulness as a preprocessing step to detect tubular structures before the actual segmentation is carried out. The former allows for an ordered growing process, which enables topological analysis. The contribution of this work is as follows. 1. A new vessel enhancement filter for tubular structures based on the Laplacian is proposed, 2. a wavefront propagation technique is proposed that prevents leaks by imposing a threshold on the maximum number of voxels that the propagating front must contain, and 3. a volumetric hole filling method is proposed to filll holes, bays, and tunnels which are caused at locations where the tubular structure assumption is violated. The proposed method reduces approximately 50% of the necessary eigenvalue calculations for vessel enhancement and prevents leaks starting at small spots, which usually occur using standard region growing. Qualitative and quantitative evaluation based on several metrics (statistical measures, dice and symmetric average surface distance) is presented.

This paper presents initial performance results for a software toolkit that implements GPU-based parallel computation of digitally reconstructed radiographs (DRRs) from volumetric imaging data for 2D-3D registration. The computational parallelism is achieved using NVIDIA’s CUDA implementation of general purpose computing on the graphics processing unit. The sample volumetric imaging data shown here is from CT imaging of a cadaveric foot, but the toolkit can be applied equally well to other volumetric imaging data. An efficient implementation requires launching hundreds of simultaneous, independent computational threads and fast thread access to the global memory where they need to read and write data. We have implemented fast DRR generation by launching a computational thread for each pixel in the image, and achieve efficient memory access by using 3D texture memory to store the volumetric data and constant memory to store global information such as intensifier coordinates. The Thrust software library was used to store individual bone DRRs, which enables efficient memory transfer and use of built-in device operators during image compositing and similarity quantification. By storing individual DRRs, the toolkit can support independent kinematics for up to 32 segmented objects. We show that the algorithm scales with the number of processors and compare timings for three commercially available GPUs. Here we present our initial fast DRR computations to demonstrate that the toolkit can produce useful results for a full 160 × 339 × 439 stack of floating point density data on a high resolution 1152 × 896 pixel screen in 1.3 ms and on a 512 × 512 pixel screen in less than 0.6 ms.

Mammography is the most commonly used imaging modality in breast cancer screening and diagnosis. The analysis of 2D mammographic images can be difficult due to the projective nature of the imaging technique and poor contrast between tumorous and healthy fibro-glandular tissue. Contrast-enhanced magnetic resonance imaging (MRI) can overcome these disadvantages by providing a 3D dataset of the breast. The detection of corresponding image structures is challenging due to large breast deformations during the image acquisition. We present a method for analyzing 2D/3D intra-individual correspondences between mammography and MRI datasets. Therefore, an ICP-based B-spline registration is used to approximate the breast deformation differences. The resulting deformed MR image is projected onto the 2D plane to enable a comparison with the 2D mammogram. A first evaluation based on six mammograms revealed an average accuracy of 4.87 mm. In contrast to previous FEM-based approaches, we propose a fast and easy to implement 3D/3D-registration, for simulating the mammographic breast compression.

In orthodontics, a common practice used to diagnose and plan the treatment is the dental cast. After digitization by a CT-scan or a laser scanner, the obtained 3D surface models can feed orthodontics numerical tools for computer-aided diagnosis and treatment planning. One of the pre-processing critical steps is the 3D registration of dental arches to obtain the occlusion of these numerical models. For this task, we propose a vision based method to automatically compute the registration based on photos of patient mouth. From a set of matched singular points between two photos and the dental 3D models, the rigid transformation to apply to the mandible to be in contact with the maxillary may be computed by minimizing the reprojection errors. During a precedent study, we established the feasibility of this visual registration approach with a manual selection of singular points. This paper addresses the issue of automatic point detection. Based on a priori knowledge, histogram thresholding and edge detection are used to extract specific points in 2D images. Concurrently, curvatures information detects 3D corresponding points. To improve the quality of the final registration, we also introduce a combined optimization of the projection matrix with the 2D/3D point positions. These new developments are evaluated on real data by considering the reprojection errors and the deviation angles after registration in respect to the manual reference occlusion realized by a specialist.

For quantitative MRI techniques, such as T1, T2 mapping and Diffusion Tensor Imaging (DTI), a model has to be fit to several MR images that are acquired with suitably chosen different acquisition settings. The most efficient estimator to retrieve the parameters is the Maximum Likelihood (ML) estimator. However, the standard ML estimator is biased for finite sample sizes. In this paper we derive a bias correction formula for magnitude MR images. This correction is applied in two different simulation experiments, a T2 mapping experiment and a DTI experiment. We show that the correction formula successfully removes the bias. As the correction is performed as post-processing, it is possible to retrospectively correct the results of previous quantitative experiments. With this procedure more accurate quantitative values can be obtained from quantitative MR acquisitions.

This paper presents a method for iterative minimization of combined residual and prediction error for near-lossless compression of medical computed tomography acquisitions using pixel-wise least-squares prediction. While most other lossy state-of-the-art image compression systems like JPEG 2000 make use of transform-based coding, in lossless coding higher compression ratios can be achieved with plain predictive algorithms like JPEG-LS because of their non-linear data adaptive energy reduction. Yet, applying these algorithms in lossy coding, simple quantization usually leads to error propagation and therefore serious quality loss or rate increase, as prediction accuracy of a pixel value and thus data rate depends on the previously reconstructed image region. The proposed minimization approach modifies the original image to be coded in a way such that the edge-directed prediction method from literature may achieve better predictions while introducing only a minimum amount of distortion. Compared to transform-based coding methods, the distortion introduced by the proposed scheme mostly consists in noise reduction instead of blurring or the introduction of artificial structures. The method also prevents error propagation due to the consideration of all pixel dependencies of the prediction. It is shown that, combined with a context-adaptive arithmetic coder, in high-fidelity coding (i. e., PSNR higher than 55 dB) the proposed method can achieve higher compression ratios than the transform-based approaches JPEG 2000, H.264/AVC, and HEVC intra coding.

In this work we propose a novel method for brain tumor segmentation in MRI by adapting the sparse optimization techniques. The core of the method lies in the subspace decomposition of the tissue feature space constituted by the brain MR images. The tumor-grown MRI slices can be viewed as a corrupted observation, which therefore can be decomposed into two components: the low-rank normal brain tissue structures and the sparse corruption/error that is due to the developed tumor. Through performing rank decomposition the corruption/error can be spotted out, thus giving rise to an initial segmentation of tumor. Our method requires no model learning. Experiments are performed on a data set of 12 subjects and the segmentation agreement is 0.86 in terms of the Dice’s similarity coefficient in comparison with the manual segmentation that is performed by a 15-year experienced radiologist. The proposed method represents an efficient mode for brain tumor segmentation that may be potentially incorporated in automated or semi-automatic segmentation systems in the clinical workflow.

Segmentation of blood vessels from magnetic resonance angiography (MRA) or computed tomography angiography (CTA) images is a complex process that usually takes a lot of computational resources. Also, most vascular segmentation and detection algorithms do not work properly due to the wide architectural variability of the blood vessels. Thus, the construction of convincing synthetic vascular trees makes it possible to validate new segmentation methodologies. In this work, an extension to the traditional Lindenmayer system (L-system) that generates synthetic 3D blood vessels by adding stochastic rules and parameters to the grammar is proposed. Towards this aim, we implement a parser and a generator of L-systems whose grammars simulate natural features of real vessels such as the bifurcation angle, average length and diameter, as well as vascular anomalies, such as aneurysms and stenoses. The resulting expressions are then used to create synthetic vessel images that mimic MRA and CTA images. In addition, this methodology allows for vessel growth to be limited by arbitrary 3D surfaces, and the vessel intensity profille can be tailored to match real angiographic intensities.

This paper presents a patch based method to normalize temporal intensities from longitudinal brain magnetic resonance (MR) images. Longitudinal intensity normalization is relevant for subsequent processing, such as segmentation, so that rates of change of tissue volumes, cortical thickness, or shapes of brain structures becomes stable and smooth over time. Instead of using intensities at each voxel, we use patches as image features as a patch encodes neighborhood information of the center voxel. Once all the time-points of a longitudinal dataset are registered, the longitudinal intensity change at each patch is assumed to follow an auto-regressive (AR(1)) process. An estimate of the normalized intensities of a patch at every time-point are generated from a hidden Markov model, where the hidden states are the unobserved normalized patches and the outputs are the observed patches. A validation study on a phantom dataset shows good segmentation overlap with the truth, and an experiment with real data shows more stable rates of change for tissue volumes with the temporal normalization than without.

Volumetric measurements of neonatal brain tissue classes have been suggested as an indicator of long-term neurodevelopmental performance. To obtain these measurements, accurate brain tissue segmentation is needed. We propose a novel method for automatic segmentation of cortical grey matter (CoGM), unmyelinated white matter (UWM), myelinated white matter (MWM), basal ganglia and thalami, brainstem, cerebellum, ventricles, and cerebrospinal fluid in the extracerebral space (CSF) in MRI scans of the brain in preterm infants. For this project, seven preterm born infants, scanned at term equivalent age were used. Axial T1- and T2- weighted scans were acquired with 3T MRI scanner. The automatic segmentation was performed in three subsequent stages where each tissue was labeled. First, a multi-atlas-based segmentation (MAS) was employed to obtain localized, subject specific spatially varying priors for each tissue. Next, based on these priors, two-class classification with k-nearest neighbor (kNN) classifier was performed to obtain the segmentation of each tissue type separately. Last, to refine the final result, and to achieve the segmentation along the tissue boundaries, a multiclass naive Bayes classifier was employed. The results were evaluated against the manually set reference standard and quantified in terms of Dice coefficient (DC) and modified Hausdorff distance (MHD), defined as 95th-percentile of the Hausdorff distance. On average, the method achieved the following DCs: 0.87 for CoGM, 0.91 for UWM, 0.60 for MWM, 0.93 for basal ganglia and thalami, 0.87 for brainstem, 0.94 for cerebellum, 0.86 for ventricles, 0.82 for CSF. The obtained average MHDs were 0.48 mm, 0.44 mm, 3.09 mm, 0.39 mm, 0.62 mm, 0.35 mm, 1.75 mm, 1.13 mm, for each tissue, respectively. The proposed methods achieved high segmentation accuracy for all tissues, except for MWM, and it provides a tool for quantification of brain tissue volumes in axial MRI scans of preterm born infants.

Labeling or segmentation of structures of interest on medical images plays an essential role in both clinical and scientific understanding of the biological etiology, progression, and recurrence of pathological disorders. Here, we focus on the optic nerve, a structure that plays a critical role in many devastating pathological conditions – including glaucoma, ischemic neuropathy, optic neuritis and multiple-sclerosis. Ideally, existing fully automated procedures would result in accurate and robust segmentation of the optic nerve anatomy. However, current segmentation procedures often require manual intervention due to anatomical and imaging variability. Herein, we propose a framework for robust and fully-automated segmentation of the optic nerve anatomy. First, we provide a robust registration procedure that results in consistent registrations, despite highly varying data in terms of voxel resolution and image field-of-view. Additionally, we demonstrate the efficacy of a recently proposed non-local label fusion algorithm that accounts for small scale errors in registration correspondence. On a dataset consisting of 31 highly varying computed tomography (CT) images of the human brain, we demonstrate that the proposed framework consistently results in accurate segmentations. In particular, we show (1) that the proposed registration procedure results in robust registrations of the optic nerve anatomy, and (2) that the non-local statistical fusion algorithm significantly outperforms several of the state-of-the-art label fusion algorithms.