Biomedical Image Registration

Image registration plays an important role within pulmonary image analysis. Accurate registration is critical to post-analysis of lung mechanical properties. To improve registration accuracy, we utilize the rich information of vessel locations and shapes, and introduce a new similarity criterion, sum of squared vesselness measure difference (SSVMD). This metric is added to three existing intensity-based similarity criteria for nonrigid lung CT image registration to show its ability in improving matching accuracy. The registration accuracy is assessed by landmark error calculation and distance map visualization on vascular tree. The average landmark errors are reduced by over 20% and are within 0.7 mm after adding SSVMD constraint to three existing intensity-based similarity metrics. Visual inspection shows matching accuracy improvements in the lung regions near the thoracic cage and near the diaphragm. Experiments also show this vesselness constraint makes the Jacobian map of transformations physiologically more plausible and reliable.

A novel method for the deformable image registration of follow-up breast magnetic resonance (MR) images is proposed, aimed at an automatic synchronization of temporal images. To compensate potentially large breast deformations and differences among device coordinates, an initial linear alignment of each individual breast, a combination of both transformations using thin-plate splines, as well as a subsequent linear-elastic registration are performed in sequence. Complementary to algorithmic details, an overview of modality-specific factors influencing follow-up registration accuracy is given. The proposed method was evaluated on 20 clinical datasets annotated with landmarks by an expert radiologist. Despite large variations among the MR images, accuracy of the method was sufficient to allow spatial synchronization, with remaining target registration errors of < 32%. Concluding, potential enhancements to further increase robustness and accuracy are discussed.

This work presents a complete processing-chain for a 3D-reconstruction of Basal Cell Carcinoma (BCC). BCC is the most common malignant skin cancer with a high risk of local recurrence after insufficient treatment. Therefore, we have focused on the development of an automated image-processing chain for 3D-reconstruction of BCC using large histological serial sections. We introduce a novel kind of image-processing chain (core component: non-linear image registration) which is optimised for the diffuse nature of BCC.

For full-automatic delineation of the tumour within the tissue we apply a fuzzy c-means segmentation method, which does not calculate a hard segmentation decision but class membership probabilities. This feature moves the binary decision tumorous vs. non-tumorous to the end of the processing chain, and it ensures smooth gradients which are needed for a consistent registration.

We used a multi-grid form of the nonlinear registration effectively suppressing registration runs into local minima (possibly caused by diffuse nature of the tumour). To register the stack of images this method is applied in a new way to reduce a global drift of the image stack while registration.

Our method was successfully applied in a proof-of-principle study for automated tissue volume reconstruction followed by a quantitative tumour growth analysis.

The registration of 3D vasculature to 2D projections is the key for providing advanced systems for image-based navigation and guidance. In areas with non-rigid patient motion, however, it is very difficult to accurately perform the registration if only one 2D view is available.

We propose a method for deformable registration of a 3D vascular model extracted from an angiographic scan to a single 2D Digitally Subtracted Angiogram (DSA). Different to existing approaches, our method does not require a segmentation of 2D vasculature. In consequence, our method can be used without manual interaction during medical treatment.

Formulated as an energy minimization problem, our approach combines a novel data term with the length regularization proposed in [1] which removes the ill-posedness of this monocular scenario. Besides attracting projected 3D centerline points to locations with high vessel probability the proposed data term ensures an injective projection of the centerline points.

Due to our novel image-based data term, we achieve a considerable gain in performance compared to feature-based approaches.

Accuracy, robustness to outliers, as well as performance issues are analyzed through tests on synthetic and real data within a controlled environment.

We introduce a method for volumetric cardiac motion analysis using variational optical flow computation involving the prior with the fractional order differentiations. The order of the differentiation of the prior controls the continuity class of the solution. Fractional differentiations is a typical tool for edge detection of images. As a sequel of image analysis by fractional differentiation, we apply the theory of fractional differentiation to a temporal image sequence analysis. Using the fractional order differentiations, we can estimate the orders of local continuities of optical flow vectors. Therefore, we can obtain the optical flow vector with the optimal continuity at each point.

Images from high-resolution scanning ophthalmic instruments are significantly distorted due to eye movement. Accurate image registration is required to successfully image subjects who are unable to fixate due to retinal conditions. Moreover, all scanning ophthalmic imaging modalities using adaptive optics will benefit from image registration, even in subjects with good fixation and anaesthetized animals. Transformation functions used to map two images could in principle be very complex. Here, we show that when the scanning in ophthalmic instruments is sufficiently fast with respect to the speed of involuntary eye movement, these mapping functions become the addition of a linear term and a single variable function. Then, based on experimental data on eye movement amplitude and speed of the fixating eye, minimum sampling frequencies for these instruments are discussed. Finally, a simple method for estimating the image transformation functions by taking advantage of the finite bandwidth of the motion signals is presented.

This paper is focused on multimodal and multidimensional image data registration: the three-dimensional retinal optical coherence tomographic (OCT) data and two-dimensional color images of fundus. The registration of these two modalities is not common in retinal image processing, but it might help to remove the moving artifacts in OCT and correct the true positions of acquired OCT scans on retina. The proposed framework consists of three steps: global dataset pre-registration, preprocessing and OCT to fundus image registration. Two alternating registration criteria are used in the main step due to changing spatial image properties. Three-parametric spatial transformation (shift and scale) for each OCT scan and exhaustive search is used in this preliminarily work. The achieved results are presented on several examples.

Voxel-based analysis, either whole-brain or tract-specific, is a widely used approach for localizing white matter (WM) differences across populations using diffusion tensor imaging (DTI). A prerequisite to this approach is to spatially normalize all the subjects to a common template. The accuracy of spatial normalization can be improved by using a population-specific template that is, morphologically, most similar to the subjects in the population of interest. Here, we report the development of a population-specific DTI template for the elderly using the publicly available IXI brain database. The template captures the average shape and diffusion properties of the aging population and contains segmentations of major WM fasciculi parcellated via fiber tractography. Furthermore, the segmentations are modeled using surface-based representation to support the tract-specific analysis recently proposed by Yushkevich et al. The utility of the template is demonstrated in an examination of WM changes in Amyotrophic Lateral Sclerosis.

In many problems involving multiple image analysis, an image registration step is required. One such problem appears in brain tumor imaging, where baseline and follow-up image volumes from a tumor patient are often to-be compared. Nature of the registration for a change detection problem in brain tumor growth analysis is usually rigid or affine. Contrast enhanced T1-weighted MR images are widely used in clinical practice for monitoring brain tumors. Over this modality, contours of the active tumor cells and whole tumor borders and margins are visually enhanced. In this study, a new technique to register serial contrast enhanced T1 weighted MR images is presented. The proposed fully-automated method is based on five anatomical landmarks: eye balls, nose, confluence of sagittal sinus, and apex of superior sagittal sinus. After extraction of anatomical landmarks from fixed and moving volumes, an affine transformation is estimated by minimizing the sum of squared distances between the landmark coordinates. Final result is refined with a surface registration, which is based on head masks confined to the surface of the scalp, as well as to a plane constructed from three of the extracted features. The overall registration is not intensity based, and it depends only on the invariant structures. Validation studies using both synthetically transformed MRI data, and real MRI scans, which included several markers over the head of the patient were performed. In addition, comparison studies against manual landmarks marked by a radiologist, as well as against the results obtained from a typical mutual information based method were carried out to demonstrate the effectiveness of the proposed method.

Elastic deformation models are frequently used when solving non-rigid registration problems that are associated with neurosurgical image guidance, however, establishing precise values for the material parameters of brain tissue remains challenging. In this work we include elastography in the registration process by formulating these parameters as unknown random variables with associated priors that may be broad or sharp, depending on the situation.

A Bayesian registration model is introduced where the deformation probability is formulated by way of Boltzmann’s equation and the linear elastic deformation and similarity energies. The full joint posterior on deformation and elastic random variables is characterized with a Markov Chain Monte Carlo method and can provide useful information beyond the usual “point estimates”; e.g. deformation uncertainty. Hard deformation constraints are easily accommodated in this framework which allows us to constrain the deformation of the brain to the intra-cranial space.

We describe preliminary experiments with synthetic 3D brain images for which ground truth is known for the elastic and deformation parameters. We compare a model with separate elastic parameters for three compartments (white matter, gray matter, and CSF), to a single compartment model, and show convergence, improved deformation estimates for the three compartment model and that plausible posteriors on the elastic parameters are obtained from the elastography process.

Validation of respiratory motion estimation is indispensable for a variety of clinical applications. For CT lung registration, current approaches employ manually defined landmark sets or contours and compute a target registration error (TRE) to quantify registration accuracy. Preferably, the landmark set is well-dispersed to reflect for lung anatomy with its varying degrees of stiffness. A recent comparison study, however, revealed that the TRE is not sufficient for functional lung analysis.

On the basis of a compressible CT phantom functional lung analysis is addressed. Non-plausible expansion patterns as they occur for CT lung data are analyzed. Motivated by the relation of Hounsfield value and local volume change, local stiffness is incorporated into registration such that an improved functional lung analysis is achieved.

Longitudinal brain image studies quantify the changes happening over time. Jacobian maps, which characterize the volume change, are based on non-rigid registration techniques and do not always appear to be clinically plausible. In particular, extreme values of volume change are not expected to be seen. The Free-Form Deformation (FFD) algorithm suffers from this drawback. Different penalty terms have been proposed in the past. We present in this paper a regularisation of the B-Spline displacements using nonlinear elasticity. Our work links a finite element method with pseudo-forces derived from a similarity measure. The presented method has been evaluated on longitudinal T1-weighted MR images of Huntington’s disease subjects and controls. Multiple time point consistency, the Jacobian map homogeneity and statistical power for group separation have been used. Our new method performs better than the classical FFD, while keeping similar registration accuracy.

This paper describes the functionality and use of the Non-rigid Image Registration Evaluation Program (NIREP) that was developed to make qualitative and quantitative performance comparisons between one or more image registration algorithms. Registration performance is evaluated using common evaluation databases. An evaluation database consists of groups of registered medical images (e.g., one or more MRI modalities, CT, etc.) and annotations (e.g., segmentations, landmarks, contours, etc.) identified by their common image coordinate system. Prior to analysis with NIREP, each algorithm is used to generate pair-wise correspondence maps/transformations between image coordinate systems. NIREP has a highly customizable graphical user interface for displaying images, transformations, segmentations, overlays, differences between images, and differences between transformations. Evaluation statistics built into NIREP are used to compute quantitative algorithm performance reports that include region of interest overlap, intensity variance of images mapped to a reference coordinate system, inverse consistency error and transitivity error.

Before an image registration method can be used in the medical theater a rigorous performance assessment of the registration method must be performed. In this paper, a new image database with a reference-based standardized evaluation methodology for objective evaluation and comparison of 3D/2D registration methods has been introduced. CT images of a female from the Visible Human Project^® were used and 15 sub-volumes each containing one of the vertebrae T3-T12 and L1-L5, and the pelvis were defined. Three pairs of lateral and anterior-posterior 2D fluoroscopic X-ray images were rendered from the CT data. Ray-casting algorithm with an energy conversion function was used to generate realistic fluoroscopic-like DRR images. Furthermore, outliers similar to medical intervention tools were also simulated on the 2D images. The assessment protocol to evaluate four criteria: accuracy, reliability, robustness and algorithm complexity, was defined. The proposed image database with the standardized evaluation methodology comprising ground truth registrations, displacements from the ground truth and target points is available upon request from the authors.

We propose to characterize deformable registration methods in a unified way, based on their parametrization. In contrast to traditional classifications, we do not apply this characterization only to standard “parametric” methods such as B-Spline Free-form deformations, but we explicitly include elastic and fluid-type “non-parametric” methods, such as the classic variational approach, and the fluid demons method. To this end, we consider parametrizations by linear combinations of arbitrary basis functions. While for the variational approach we simply utilize piecewise linear bases, for the fluid demons method we provide a new interpretation by showing that it can be seen as inherently parametrized by densely located Gaussian basis functions. Furthermore, we show that the semi-implicit discretization of the variational approach can be seen as steepest descent, with a displacement parametrized by densely located bases, based on Green’s functions corresponding to the regularization. This provides a further connection to the demons approaches. The proposed characterization is widely applicable and provides a simple and intuitive way of relating some of the arguably most commonly used methods to each other.

We propose a reliability measure that identifies informative image cues useful for registration, and present a novel, data-driven approach to spatially adapt regularization to the local image content via use of the proposed measure. We illustrate the generality of this adaptive regularization approach within a powerful discrete optimization framework and present various ways to construct a spatially varying regularization weight based on the proposed measure. We evaluate our approach within the registration process using synthetic experiments and demonstrate its utility in real applications. As our results demonstrate, our approach yielded higher registration accuracy than non-adaptive approaches and the proposed reliability measure performed robustly even in the presences of noise and intensity inhomogenity.

In this paper we present two fine and coarse approaches for the efficient registration of 3D medical images using the framework of Large Deformation Diffeomorphic Metric Mapping (LDDMM). This formalism has several important advantages since it allows large, smooth and invertible deformations and has interesting statistical properties. We first highlight the influence of the smoothing kernel in the LDDMM framework. We then show why approaches taking into account several scales simultaneously should be used for the registration of complex shapes, such as those treated in medical imaging. We then present our fine and coarse approaches and apply them to the registration of binary images as well as the longitudinal estimation of the early brain growth in preterm MR images.

Diffusion tensor imaging provides information about deep white matter anatomy that structural magnetic resonance images typically fail to resolve. Non-linear registration of diffusion tensor images, for which a few methods already exist, allows us to capture the deformations of these structures that would otherwise go unobserved. Here, we build on an existing method for diffeomorphic registration of diffusion tensor images, so that it fully incorporates the useful log-domain parameterization of diffeomorphisms. Initially, this allows us to easily include a registration symmetry constraint that is highly desirable for pair-wise registration. More importantly, the parameterization allows simple and proper calculation of statistics on the transformations obtained. We show that the symmetric log-domain method exhibits the most preferable trade-off between image correspondence and deformation smoothness on real data and also achieves the best recovery of synthetic warps.

In this paper, we present a method for coronary artery motion tracking in 4D cardiac CT angiogram data sets. The proposed method allows the construction of patient-specific 4D coronary motion model from pre-operative CTA which can be used for guiding totally endoscopic coronary artery bypass surgery (TECAB). The proposed approach consists of three steps: Firstly, the coronary arteries are extracted in the end-diastolic time frame using a minimal cost path approach. To achieve this, the start and end points of the coronaries are identified interactively and the minimal cost path between the start and end points is computed using A* graph search algorithm. Secondly, the cardiac motion is estimated throughout the cardiac cycle by using a non-rigid image registration technique based on a free-form B-spline transformation model and maximization of normalized mutual information. Finally, coronary arteries are tracked automatically through all other phases of the cardiac cycle. This is estimated by deforming the extracted coronaries at end-diastole to all other time frames according the motion field acquired in second step. The estimated coronary centerlines are then refined by template matching algorithm to improve the accuracy. We compare the proposed approach with two alternative approaches: The first approach is based on the minimal cost path extraction of the coronaries with start and end points manually identified in each time frame while the second approach is based on propagating the extracted coronaries from the end-diastolic time frame to other time frames using image-based non-rigid registration only. Our results show that the proposed approach performs more robustly than the non-rigid registration based method and that the resulting motion model is comparable to the motion model constructed from semi-automatic extractions of the coronaries in all time frames.

Motion models have been widely applied as a solution to the problem of organ motion in both image acquisition and image guided interventions. The traditional approach to constructing motion models from dynamic images involves first coregistering the images to produce estimates of motion parameters, and then modelling the variation of these parameters as functions of a surrogate value or values. Errors in this approach can result from inaccuracies in the image registrations and in the modelling process. In this paper we describe an approach in which the registrations of all images and the modelling process are performed simultaneously. Using numerical phantom data and 21 dynamic magnetic resonance imaging (MRI) datasets acquired from 7 volunteers and 7 patients, we demonstrate that our new technique results in an average reduction in motion model errors of 11.5% for the phantom experiments and 1.8% for the MRI experiments. This approach has the potential to improve the accuracy of motion estimates for a range of applications.

Radio-frequency catheter ablation (RFCA) has become an accepted treatment option for atrial fibrillation (Afib). RFCA of Afib involves isolation of the pulmonary veins under X-ray guidance. For easier navigation, two-dimensional X-ray imaging may take advantage of overlay images derived from static pre-operative 3-D data set to add anatomical details which, otherwise, would not be visible under X-ray. Unfortunately, respiratory and cardiac motion may impair the utility of static overlay images for catheter navigation. We developed a system for image-based 2-D motion estimation and compensation as a solution to this problem. It is based on 2-D catheter tracking facilitated by model-based registration of an ellipse-shaped model to fluorosocpic images. A mono-plane or a bi-plane X-ray C-arm system can be used. In the first step of the method, a 2-D model of the catheter device is computed. Respiratory and cardiac motion at the site of ablation is then estimated by tracking the catheter device in fluoroscopic images. The cost function of the registration step is based on the average distance of the model to the segmented circumferential mapping catheter using a distance map. In our experiments, the circumferential catheter was successfully tracked in 688 fluoroscopic images with an average 2-D tracking error of 0.59 mm ± 0.25 mm. Our presented method achieves a tracking rate of 10 frames-per-second.

We propose a new nonrigid registration method based on a unified framework of encoding spatial information in entropy measures. The encoding of spatial information improves nonrigid registration against the problems caused by intensity distortion where the registration using traditional mutual information (MI) is challenged. Using this encoding framework, we derive the new registration method, spatial information encoded mutual information (SIEMI). SIEMI registration has a similar computation complexity as the registration using traditional MI measures, but works significantly better in the nonrigid cases. We validated the registration method using brain MRI and dynamic contrast enhanced MRI of the liver. The results showed that the proposed method performed significantly better than the normalized mutual information registration.

Mutual information (MI) was introduced for use in multimodal image registration over a decade ago [1,2,3,4]. The MI between two images is based on their marginal and joint/conditional entropies. The most common versions of entropy used to compute MI are the Shannon and differential entropies; however, many other definitions of entropy have been proposed as competitors. In this article, we show how to construct normalized versions of MI using any of these definitions of entropy. The resulting similarity measures are analogous to normalized mutual information (NMI), entropy correlation coefficient (ECC), and symmetric uncertainty (SU), which have all been shown to be superior to MI in a variety of situations. We use publicly available CT, PET, and MR brain images with known ground truth transformations to evaluate the performance of the normalized measures for rigid multimodal registration. Results show that for a number of different definitions of entropy, the proposed normalized versions of mutual information provide a statistically significant improvement in target registration error (TRE) over the non-normalized versions.

We propose Mooney-Rivlin (MR) nonlinear elasticity of hyperelastic materials and numerical algorithms for image registration in the presence of landmarks and large deformation. An auxiliary variable is introduced to remove the nonlinearity in the derivatives of Euler-Lagrange equations. Comparing the MR elasticity model with the Saint Venant-Kirchhoff elasticity model (SVK), the results show that the MR model gives better matching in fewer iterations. To accelerate the slow convergence due to the lack of smoothness of the L ² gradient, we construct a Sobolev H ¹ gradient descent method [13] and take advantage of the smoothing quality of the Sobolev operator \((Id-\triangle)^{-1}\). The MR model with Sobolev H ¹ gradient descent (SGMR) improves both matching criterion and computational time substantially. We further apply the L ² and Sobolev gradient to landmark registration for multi-modal mouse brain data, and observe faster convergence and better landmark matching for the MR model with Sobolev H ¹ gradient descent.