Biomedical Image Registration

We are concerned with the segmentation of structures within the brain particularly the gyri of the cerebral cortex, but also subcortical structures from volumetric T1-weighted MRI images. A fully automatic multi-atlas registration based segmentation approach is used to label novel data. We use a standard affine registration method combined with a small deformation (non-diffeomorphic), non-linear registration method which optimises mutual information, with a cascading set of regularisation parameters. We consistently segment 138 structures in the brain, 98 in the cortex and 40 in the sub-cortex. An overall Dice score of 0.704 and a mean surface distance of 1.106 mm is achieved in leave-one-out cross validation using all available atlases. The algorithm has been evaluated on a number of different cohorts which includes patients of different ages and scanner manufacturers, and has been shown to be robust. It is shown to have comparable accuracy to other state of the art methods using the MICCAI 2012 multi-atlas challenge benchmark, but the runtime is substantially less.

B-spline interpolation is a widely used interpolation technique. In the field of image registration, interpolation is necessary for transforming images to obtain a measure of (dis)similarity between the images to be aligned. When gradient-based optimization methods are used, the image gradients need to be calculated as well, which also accounts for a substantial share of computation time in registration. In this paper we propose a fast multidimensional B-spline interpolation algorithm with which both image value and gradient can be computed efficiently. We present a recursive algorithm for the interpolation which is efficiently implemented with template metaprogramming (TMP). The proposed algorithm is compared with the algorithm implemented in the Insight Toolkit (ITK), for different interpolation orders and image dimensions. Also, the effect on the computation time of a typical registration problem is evaluated. The results show that the computation time of B-spline interpolation is decreased by the proposed algorithm from a factor 4.1 for a 2D image using 1st order interpolation to a factor of 19.9 for 4D using 3rd order interpolation.

We propose a novel non-linear registration strategy which seeks an optimal deformation that maps corresponding boundaries of similar orientation. Our approach relies on a local similarity metric based on gradient orientation alignment and distance to the nearest inferred boundary and is evaluated on a reduced set of locations corresponding to inferred boundaries. The deformation model is characterized as the integration of a time-constant velocity field and optimization is performed in coarse to fine multi-level strategy with a gradient ascent technique. Our approach is computational efficient since it relies on a sparse selection of voxels corresponding to detected boundaries, yielding robust and accurate results with reduced processing times. We demonstrate quantitative results in the context of the non-linear registration of inter-patient magnetic resonance brain volumes obtained from a public dataset (CUMC12). Our proposed approach achieves a similar level of accuracy as other state-of-the-art methods but with processing times as short as 1.5 minutes. We also demonstrate preliminary qualitative results in the time-sensitive registration contexts of registering MR brain volumes to intra-operative ultrasound for improved guidance in neurosurgery.

An important research problem in image-guided radiation therapy is how to accurately register daily onboard Cone-beam CT (CBCT) images to higher quality pretreatment fan-beam CT (FBCT) images. Assuming the organ segmentations are both available on CBCT and FBCT images, methods have been proposed to use them to help the intensity-driven image registration. Due to the low contrast between soft-tissue structures exhibited in CBCT, the interobserver contouring variability (expressed as standard deviation) can be as large as 2-3 mm and varies systematically with organ, and relative location on each organ surface. Therefore the inclusion of the segmentations into registration may degrade registration accuracy. To address this issue we propose a surface assimilation method that estimates a new surface from the manual segmentation from a priori organ shape knowledge and the interobserver segmentation error. Our experiment results show the proposed method improves registration accuracy compared to previous methods.

We describe an automated method for building a statistical model of the mouse hind limb from micro-CT data, based on articulated registration. The model was initialised by hand-labelling the constituent bones and joints of a single sample. A coarse alignment of the entire model mesh to a sample mesh was followed by consecutive registration of individual bones and their descendants down a hierarchy. Transformation parameters for subsequent bones were constrained to a subset of vertices within a frustum projecting from a terminal joint of an already registered parent bone. Samples were segmented and transformed into a common coordinate frame, and a statistical shape model was constructed. The results of ten registered samples are presented, with a mean registration error of less than 40 μm (~ 3 voxels) for all samples. The shape variation amongst the samples was extracted by PCA to create a statistical shape model. Registration of the model to three unseen normal samples gives rise to a mean registration error of 5.84 μm, in contrast to 27.18 μm for three unseen arthritic samples. This may suggest that pathological bone shape changes in models of RA are detectable as departures from the model statistics.

Deformable image registration is an important step in medical image analysis. It enables an automatic labelling of anatomical structures using atlas-based segmentation, motion compensation and multi-modal fusion. The use of discrete optimisation approaches has recently attracted a lot attention for mainly two reasons. First, they are able to find an approximate global optimum of the registration cost function and can avoid false local optima. Second, they do not require a derivative of the similarity metric, which increases their flexibility. However, the necessary quantisation of the deformation space causes a very large number of degrees of freedom with a high computational complexity. To deal with this, previous work has focussed on parametric transformation models. In this work, we present an efficient non-parametric discrete registration method using a filter-based similarity cost aggregation and a decomposition of similarity and regularisation term into two convex optimisation steps. This approach enables non-parametric registration with billions of degrees of freedom with computation times of less than a minute. We apply our method to two different common medical image registration tasks, intra-patient 4D-CT lung motion estimation and inter-subject MRI brain registration for segmentation propagation. We show improvements on current state-of-the-art performance both in terms of accuracy and computation time.

B-spline based free-form deformation (FFD) is a widely used technique in nonrigid image registration. In general, a third-order B-spline function is used, because of its favorable trade-off between smoothness and computational cost. Compared with the third-order B-splines, a B-spline function with a lower order has shorter support length, which means it is computationally more attractive. However, a lower-order function is seldom used to construct the deformation field for registration since it is less smooth. In this work, we propose a randomly perturbed FFD strategy (RPFFD) which uses a lower-order B-spline FFD with a random perturbation around the original position to approximate a higher-order B-spline FFD in a stochastic fashion. For a given D-dimensional nth-order FFD, its corresponding (n − 1)th-order RPFFD has \((\frac{n}{n+1})^{D}\) times lower computational complexity. Experiments on 3D lung and brain data show that, with this lower computational complexity, the proposed RPFFD registration results in even slightly better accuracy and smoothness than the traditional higher-order FFD.

This paper presents a novel mathematical framework for representing uncertainty in large deformation diffeomorphic image registration. The Bayesian posterior distribution over the deformations aligning a moving and a fixed image is approximated via a variational formulation. A stochastic differential equation (SDE) modeling the deformations as the evolution of a time-varying velocity field leads to a prior density over deformations in the form of a Gaussian process. This permits estimating the full posterior distribution in order to represent uncertainty, in contrast to methods in which the posterior is approximated via Monte Carlo sampling or maximized in maximum a-posteriori (MAP) estimation. The framework is demonstrated in the case of landmark-based image registration, including simulated data and annotated pre and intra-operative 3D images.

We propose a robust and efficient approach for the reconstruction of the descending aorta from contiguous 3D transesophageal echocardiographic (TEE) images. It is based on an ad hoc protocol, designed to acquire ordered and partially overlapped 3D TEE datasets, followed by automated image registration that relies on this a priori knowledge. The method was validated using artificially derived misaligned images, and then applied to 14 consecutive patients. Both qualitative and quantitative results demonstrated the potential feasibility and accuracy of the proposed approach. Its clinical applicability could improve the assessment of aortic total plaque burden from 3D TEE images.

Accurate registration of retinal fundus images is vital in computer aided diagnosis of retinal diseases. This paper presents a robust registration method that makes use of the intensity as well as structural information of the retinal vasculature. In order to correct for illumination variation between images, a normalized-convolution based luminosity and contrast normalization technique is proposed. The normalized images are then aligned based on a vasculature-weighted mean squared difference (MSD) similarity metric. To increase robustness, we designed a multiresolution matching strategy coupled with a hierarchical registration model. The latter employs a deformation model with increasing complexity to estimate the parameters of a global second-order transformation model. The method was applied to combine 400 fundus images from 100 eyes, obtained from an ongoing diabetic retinopathy screening program, into 100 mosaics. Accuracy assessment by experienced clinical experts showed that 89 (out of 100) mosaics were either free of any noticeable misalignment or have a misalignment smaller than the width of the misaligned vessel.

Respiratory motion models relate the motion of the internal anatomy, which can be difficult to directly measure during image guided interventions or image acquisitions, to easily acquired respiratory surrogate signal(s), such as the motion of the skin surface. The motion models are usually built in two steps: 1) determine the motion from some prior imaging data, e.g. using image registration, 2) fit a correspondence model relating the motion to the surrogate signal(s). In this paper we present a generalized framework for combining the image registration and correspondence model fitting steps into a single optimization. Not only does this give a more theoretically efficient and robust approach to building the motion model, but it also enables the use of ‘partial’ imaging data such as individual MR slices or CBCT projections, where it is not possible to determine the full 3D motion from a single image. The framework can also incorporate motion compensated image reconstruction by iterating between model fitting and image reconstruction. This means it is possible to estimate both the motion and the motion compensated reconstruction just from the partial imaging data and a respiratory surrogate signal.

We have used a simple 2D ‘lung-like’ software phantom to demonstrate a proof of principle of our framework, for both simulated ‘thick-slice’ data and projection data, representing MR and CBCT data respectively. We have implemented the framework using a simple demons like registration algorithm and a linear correspondence model relating the motion to two surrogate signals.

Visualization and analysis of intra-operative images in imageguided radiotherapy and surgery are mainly limited to 2D X-ray imaging, which could be beneficially fused with information-rich pre-operative 3D image information by means of 3D-2D image registration. To keep the radiation dose delivered by the X-ray system low, the intra-operative imaging is usually limited to a single projection view. Registration of 3D to a single 2D image is a very challenging registration task for most of current state-of-the-art 3D-2D image registration methods. We propose a novel 3D-2D rigid registration method based on evaluation of similarity between corresponding 3D and 2D gradient covariances, which are mapped into the same space using backprojection. Normalized scalar product of covariances is computed as similarity measure. Performance of the proposed and state-of-the-art 3D-2D image registration methods was evaluated on two publicly available image datasets, one of cerebral angiograms and the other of a spine cadaver, using standardized evaluation methodology. Results showed that the proposed method outperformed the current state-of-the-art methods and achieved registration accuracy of 0.5 mm, capture range of 9 mm and success rate >80%. Considering also that GPU-enabled execution times ranged from 0.5-2.0 seconds, the proposed method has the potential to enhance with 3D information the visualization and analysis of intra-operative 2D images.

Previous amyloid positron emission tomography (PET) imaging studies have shown that Alzheimer’s disease (AD) patients exhibit higher standardised uptake value ratios (SUVRs) than healthy controls. Automatic methods for SUVR calculation in brain images are typically based on registration of PET brain data to a template, followed by computation of the mean uptake ratio in a set of regions in the template space. Resulting SUVRs will therefore have some dependence on the registration method. It is widely accepted that registration based on anatomical information provides optimal results. However, in clinical practice, good quality anatomical data may not be available and registration is often based on PET data alone. We investigate the effect of using functional and structural image information during the registration of PET volumes to a template, by comparing six registration methods: affine registration, non-linear registration using PET-driven demons, non-linear registration using magnetic resonance (MR) driven demons, and our novel joint PET-MR registration technique with three different combination weightings. Our registration method jointly registers PET-MR brain volume pairs, by combining the incremental updates computed in single-modality local correlation coefficient demons registrations. All six registration methods resulted in significantly higher mean SUVRs for diseased subjects compared to healthy subjects. Furthermore, the combined PET-MR registration method resulted in a small, but significant, increase in the mean Dice overlaps between cortical regions in the MR brain volumes and the MR template, compared with the single-modality registration methods. These results suggest that a non-linear, combined PET-MR registration method can perform at least as well as the single-modality registration methods in terms of the separation between SUVRs and Dice overlaps, and may be well suited to discriminate between populations of AD patients and healthy controls.

Co-registration of multi-modal microscopic images can integrate benefits of each modality, yet major challenges come from inherent difference between staining, distortions of specimens and various artefacts. In this paper, we propose a new interactive registration-learning method to register functional fluorescence (IF) and structural histology (HE) images in a pyramidal fashion. We synthesize HE image from the multi-channel IF image using a supervised machine learning technique and hence reduce the multi-modality registration problem into a mono-modality one, in which case the normalised cross correlation is used as the similarity measure. Unlike conventional applications of supervised learning, our classifier is not trained by ‘ground-truth’ (perfectly-registered) training dataset, as they are not available. Instead, we use a relatively noisy training dataset (affinely-registered) as an initialization and rely on the robustness of machine learning to the outliers and label updates via pyramidal deformable registration to gain better learning and predictions. In this sense, the proposed methodology has potential to be adapted in other learning problems as the manual labelling is usually imprecise and very difficult in the case of heterogeneous tissues.

Relaxation time (T ₁₀) estimation using variable flip angle sequences is a key step for pharmacokinetic (PK) analysis of tumours in DCE-MRI exams. In this study, the effects of motion within flip angle sequences on the T ₁₀ and subsequent K _trans and k _ep estimation were examined. It was found that errors in T ₁₀ estimation caused by motion had a significant impact on subsequent PK analysis. A new similarity metric, based on the T ₁₀ regression error, for groupwise motion correction of variable flip angle sequences is proposed and compared against Groupwise Normalized Mutual Information (GNMI). In rigid registration experiments on simulated data, the new metric outperformed GNMI, showing an improvement alignment of over 14% in terms of average target registration error, which is also reflected by a lower T ₁₀ estimation error. Finally, registration was applied to 46 clinical sequences to identify the average amount of motion found in this type of acquisition; this showed an estimated displacement of 0.98mm, which could lead to over 25% K _trans estimation error if motion were not corrected.

Optical microscopy imaging techniques have enabled a wide spectrum of biomedical applications. Among visualization, a quantitative analysis of tumour cell growth in lungs is of great importance. The main challenges inherently linked with such data analysis are: local contrast changes related to tissue depth, lack of clear object boundaries due to the presence of noise, and cluttering with motion artefacts due to translational shift of the specimen and non-linear lung tissue collapse. This paper aims to address these problems by introducing a novel image registration framework specifically designed to correct for motion artefacts from optical microscopy of lung tumour cells imaging. For this purpose, a previously developed modality independent neighbourhood descriptor (MIND) was adapted to cope with multiple image channels for optical microscopy data. Two versions of this novel multichannel MIND (mMIND) are here presented. The proposed registration technique estimates both rigid transformations and non-linear deformations both common in the optical microscopy volumes and time-sequences acquisition. The performance of our registration technique based on a novel multichannel image representation is demonstrated using two distinctive optical imaging data sets of lung cells: 3D volumes with translation motion artefacts only, and time-sequences with both rigid and non-linear motion artefacts. Visual inspection of the registration outcomes and reported results of the qualitative evaluation show a promising improvement compared to images without correction.

Quantitative magnetic resonance imaging (qMRI) aims to extract quantitative parameters representing tissue properties from a series of images by modeling the image acquisition process. This requires the images to be spatially aligned but, due to patient motion, anatomical structures in the consecutive images may be misaligned. In this work, we propose a groupwise non-rigid image registration method for motion compensation in qMRI. The method minimizes a dissimilarity measure based on principal component analysis (PCA), exploiting the fact that intensity changes can be described by a low-dimensional acquisition model. Using an unbiased groupwise formulation of the registration problem, there is no need to choose a reference image as in conventional pairwise approaches. The method was evaluated on three applications: modified Look-Locker inversion recovery T ₁ mapping in a porcine myocardium, black-blood variable flip-angle T ₁ mapping in the carotid artery region, and apparent diffusion coefficient (ADC) mapping in the abdomen. The method was compared to a conventional pairwise alignment that uses a mutual information similarity measure. Registration accuracy was evaluated by computing precision of the estimated parameters of the qMRI model. The results show that the proposed method performs equally well or better than an optimized pairwise approach and is therefore a suitable motion compensation method for a wide variety of qMRI applications.

In this paper we present a rigid registration approach for 4D ultrasound (US) datasets, where images are registered over time. The 3D registration approach preceding the 4D registration consists of two main steps - block-matching and outlier rejection. The outlier rejection step removes the spurious matchings’ from the block-matching module and ensures inverse consistency. For 4D registration, we perform registration of consecutive US volumes over the time series. Transformation between any two frames is estimated by taking the product of all the intermediate transforms. To avoid accumulation of error over the series of transformations, a long range feedback mechanism is proposed. A mean total registration error of 1 mm is achieved across six 4D ultrasound sequences of human liver with an execution speed of 10 Hz.

Popular intensity-based similarity measures such as (normalized) mutual information estimate statistics over the entire image, neglecting spatial relationships and local image properties. In this work, we present an adaptive multiscale image similarity measure for non-rigid registration which combines image statistics at multiple scales for a multiscale representation of regional image similarities. We validated the proposed similarity measure on simulated and clinical MR brain datasets. Results show that our approach achieves higher registration accuracy and robustness than conventional global measures or their local variations at a single scale.

Image registration is a ubiquitous technique in medical imaging. However, finding correspondences reliably between images is a difficult task since the registration problem is ill-posed and registration algorithms are only capable of finding local optima. This makes it challenging to find a suitable registration method and parametrization for a specific application. To alleviate such problems, multiple registrations can be fused which is typically done by weighted averaging, which is sensitive to outliers and can not guarantee that registrations improve. In contrast, in this work we present a Markov random field based technique which fuses registrations by explicitly minimizing local dissimilarities of deformed source and target image, while penalizing non-smooth deformations. We additionally propose a registration propagation technique which combines multiple registration hypotheses which are obtained from different indirect paths in a set of mutually registered images. Our fused registrations are experimentally shown to improve pair-wise correspondences in terms of average deformation error (ADE) and target registration error (TRE) as well as improving post-registration segmentation overlap.

Theoretically, inverse consistency in an image registration problem can be achieved by employing a diffeomorphic scheme that uses transformations parametrized by stationary velocity fields (SVF). The displacement from a given SVF, formulated as a series of self compositions of a transformation function, can be obtained by Euler integration in the time domain. However in practice, the discrete time integration produces results that are inverse inconsistent, and inverse consistency in the final solution needs to be explicitly ensured. One way of achieving this is to penalize the endpoint displacement offset obtained by evaluating a composition of the transformation with its inverse at an arbitrary point. In this paper, we propose a variation in which the displacement penalization is required only in the first composition step of the transformation thereby bringing down the computational complexity. We compare these two ways of enforcing inverse consistency by applying the registration framework on four pairs of brain magnetic resonance images. We observe that the proposed stepwise scheme maintains both precision and level of inverse consistency similar to the endpoint scheme.

Biomedical image registration faces challenging problems induced by the image acquisition process of the involved modality. A common problem is the omnipresence of noise perturbations. A low signal-to-noise ratio – like in modern dynamic imaging with short acquisition times – may lead to failure or artifacts in standard image registration techniques. A common approach to deal with noise in registration is image presmoothing, which may however result in bias or loss of information. A more reasonable alternative is to directly incorporate statistical noise models into image registration. In this work we present a general framework for registration of noise perturbed images based on maximum a-posteriori estimation. This leads to variational registration inference problems with data fidelities adapted to the noise characteristics, and yields a significant improvement in robustness under noise impact and parameter choices. Using synthetic data and a popular software phantom we compare the proposed model to conventional methods recently used in biomedical imaging and discuss its potential advantages.