PurposeApplying machine learning techniques to magnetic resonance diffusion-weighted imaging (DWI) data is challenging due to the size of individual data samples and the lack of labeled data. It is possible, though, to learn general patterns from a very limited amount of training data if we take advantage of the geometry of the DWI data. Therefore, we present a tissue classifier based on a Riemannian deep learning framework for single-shell DWI data.ApproachThe framework consists of three layers: a lifting layer that locally represents and convolves data on tangent spaces to produce a family of functions defined on the rotation groups of the tangent spaces, i.e., a (not necessarily continuous) function on a bundle of rotational functions on the manifold; a group convolution layer that convolves this function with rotation kernels to produce a family of local functions over each of the rotation groups; a projection layer using maximization to collapse this local data to form manifold based functions.ResultsExperiments show that our method achieves the performance of the same level as state-of-the-art while using way fewer parameters in the model (<10 % ). Meanwhile, we conducted a model sensitivity analysis for our method. We ran experiments using a proportion (69.2%, 53.3%, and 29.4%) of the original training set and analyzed how much data the model needs for the task. Results show that this does reduce the overall classification accuracy mildly, but it also boosts the accuracy for minority classes.ConclusionsThis work extended convolutional neural networks to Riemannian manifolds, and it shows the potential in understanding structural patterns in the brain, as well as in aiding manual data annotation.
Brain atrophy from structural magnetic resonance images (MRIs) is widely used as an imaging surrogate marker for Alzheimers disease. Their utility has been limited due to the large degree of variance and subsequently high sample size estimates. The only consistent and reasonably powerful atrophy estimation methods has been the boundary shift integral (BSI). In this paper, we first propose a tensor-based morphometry (TBM) method to measure voxel-wise atrophy that we combine with BSI. The combined model decreases the sample size estimates significantly when compared to BSI and TBM alone.
We explore a new approach for structural connectivity based segmentations of subcortical brain regions. Connectivity based segmentations are usually based on fibre connections from a seed region to predefined target regions. We present a method for finding significantly connected voxels based on the distribution of connection strengths. Paths from seed voxels to all voxels in a target region are obtained from a shortest-path tractography. For each seed voxel we approximate the distribution with a histogram of path scores. We hypothesise that the majority of estimated connections are false-positives and that their connection strength is distributed differently from true-positive connections. Therefore, an empirical null-distribution is defined for each target region as the average normalized histogram over all voxels in the seed region. Single histograms are then tested against the corresponding null-distribution and significance is determined using the false discovery rate (FDR). Segmentations are based on significantly connected voxels and their FDR. In this work we focus on the thalamus and the target regions were chosen by dividing the cortex into a prefrontal/temporal zone, motor zone, somatosensory zone and a parieto-occipital zone. The obtained segmentations consistently show a sparse number of significantly connected voxels that are located near the surface of the anterior thalamus over a population of 38 subjects.
Obtaining regional volume changes from a deformation field is more precise when using simplex counting (SC) compared with Jacobian integration (JI) due to the numerics involved in the latter. Although SC has been proposed before, numerical properties underpinning the method and a thorough evaluation of the method against JI is missing in the literature. The contributions of this paper are: (a) we propose surface propagation (SP)—a simplification to SC that significantly reduces its computational complexity; (b) we will derive the orders of approximation of SP which can also be extended to SC. In the experiments, we will begin by empirically showing that SP is indeed nearly identical to SC, and that both methods are more stable than JI in presence of moderate to large deformation noise. Since SC and SP are identical, we consider SP as a representative of both the methods for a practical evaluation against JI. In a real application on Alzheimer’s disease neuroimaging initiative data, we show the following: (a) SP produces whole brain and medial temporal lobe atrophy numbers that are significantly better than JI at separating between normal controls and Alzheimer’s disease patients; (b) SP produces disease group atrophy differences comparable to or better than those obtained using FreeSurfer, demonstrating the validity of the obtained clinical results. Finally, in a reproducibility study, we show that the voxel-wise application of SP yields significantly lower variance when compared to JI.
A challenge when using current magnetic resonance (MR)-based attenuation correction in positron emission tomography/MR imaging (PET/MRI) is that the MRIs can have a signal void around the dental fillings that is segmented as artificial air-regions in the attenuation map. For artifacts connected to the background, we propose an extension to an existing active contour algorithm to delineate the outer contour using the nonattenuation corrected PET image and the original attenuation map. We propose a combination of two different methods for differentiating the artifacts within the body from the anatomical air-regions by first using a template of artifact regions, and second, representing the artifact regions with a combination of active shape models and k-nearest-neighbors. The accuracy of the combined method has been evaluated using 25 F18-fluorodeoxyglucose PET/MR patients. Results showed that the approach was able to correct an average of 97±3% of the artifact areas.
Interpolating kernels are crucial to solving a stationary velocity field (SVF) based image registration problem. This is because, velocity fields need to be computed in non-integer locations during integration. The regularity in the solution to the SVF registration problem is controlled by the regularization term. In a variational formulation, this term is traditionally expressed as a squared norm which is a scalar inner product of the interpolating kernels parameterizing the velocity fields. The minimization of this term using the standard spline interpolation kernels (linear or cubic) is only approximative because of the lack of a compatible norm. In this paper, we propose to replace such interpolants with a norm-minimizing interpolant - the Wendland kernel which has the same computational simplicity like B-Splines. An application on the Alzheimer's disease neuroimaging initiative showed that Wendland SVF based measures separate (Alzheimer's disease v/s normal controls) better than both B-Spline SVFs (p<0.05 in amygdala) and B-Spline freeform deformation (p<0.05 in amygdala and cortical gray matter).
Treatment of cervical cancer, one of the three most commonly diagnosed cancers worldwide, often relies on delineations of the tumour and metastases based on PET imaging using the contrast agent 18F-Fluorodeoxyglucose (FDG). We present a robust automatic algorithm for segmenting the gross tumour volume (GTV) and metastatic lymph nodes in such images. As the cervix is located next to the bladder and FDG is washed out through the urine, the PET-positive GTV and the bladder cannot be easily separated. Our processing pipeline starts with a histogram-based region of interest detection followed by level set segmentation. After that, morphological image operations combined with clustering, region growing, and nearest neighbour labelling allow to remove the bladder and to identify the tumour and metastatic lymph nodes. The proposed method was applied to 125 patients and no failure could be detected by visual inspection. We compared our segmentations with results from manual delineations of corresponding MR and CT images, showing that the detected GTV lays at least 97.5% within the MR/CT delineations. We conclude that the algorithm has a very high potential for substituting the tedious manual delineation of PET positive areas.
MRI-determined measurement of synovial inflammation (synovitis) from hand MRIs has recently gained
considerable popularity as a secondary marker in rheumatoid arthritis (RA) clinical trials. The currently
accepted scoring systems are, however, purely semi-quantitative and rely on assessment from a trained
radiologist. We propose a novel, fully automatic technique for quantitative wrist synovitis measurement
from two MRIs acquired before and after contrast agent injection. The technique estimates the volume of
the synovial inflammation in three steps. First, the wrist synovial membrane is segmented using multi-atlas
B-spline based freeform registration. Second, positioning differences between the pre- and post-contrast
acquisitions are corrected by rigid registration. Finally, wrist synovitis is quantified from the difference
between the pre- and post-contrast sequences in the region of the segmented synovium. We evaluate the
proposed technique on a data set of nineteen patients with acquisitions at two time points in a leave-one-patient-out fashion. Our experiments show that we are able to perform synovitis measurement with good
correlation to manual semi-quantitative RAMRIS scores for both static (r=0.84) and longitudinal (r=0.87)
scoring. These results compare favorably to the RAMRIS inter-observer variability.
Various approaches have been proposed for segmentation of cardiac MRI. An accurate segmentation of the
myocardium and ventricles is essential to determine parameters of interest for the function of the heart, such as
the ejection fraction. One problem with MRI is the poor resolution in one dimension.
A 3D registration algorithm will typically use a trilinear interpolation of intensities to determine the intensity
of a deformed template image. Due to the poor resolution across slices, such linear approximation is highly
inaccurate since the assumption of smooth underlying intensities is violated. Registration-based interpolation
is based on 2D registrations between adjacent slices and is independent of segmentations. Hence, rather than
assuming smoothness in intensity, the assumption is that the anatomy is consistent across slices. The basis for
the proposed approach is the set of 2D registrations between each pair of slices, both ways. The intensity of a
new slice is then weighted by (i) the deformation functions and (ii) the intensities in the warped images. Unlike
the approach by Penney et al. 2004, this approach takes into account deformation both ways, which gives more
robustness where correspondence between slices is poor.
We demonstrate the approach on a toy example and on a set of cardiac CINE MRI. Qualitative inspection reveals
that the proposed approach provides a more convincing transition between slices than images obtained by linear
interpolation. A quantitative validation reveals significantly lower reconstruction errors than both linear and
registration-based interpolation based on one-way registrations.
We address the problem of intra-subject registration for change detection. The goal is to separate stationary and changing
subsets to be able to robustly perform rigid registration on the stationary subsets and thus improve the subsequent change
detection. An iterative approach using a hybrid of parametric and non-parametric statistics is presented. The method
uses non-parametric clustering and large scale hypothesis testing with estimation of the empirical null hypothesis. The
method is successfully applied to 3D surface scans of human ear impressions containing true changes as well as data with
synthesized changes. It is shown that the method improves registration and is capable of reducing the difference between
registration using different norms.
This work describes a non-rigid registration method for open 2D manifold embedded in 3D Euclidian space. The
method is based on difference of distance maps and grid based warps interpolated by splines constrained in such
a way that the deformation field is diffeomorphic. We then create a dense surface to surface correspondence using
angle weighted normals and ray tracing. The implementation using a derivation of the inverse compositional
algorithm for optimization of computational speed is described. The results are evaluated as a shape model
showing the principal modes of variation.
We evaluate a novel method for fully automated rigid registration of 2D manifolds in 3D space based on distance
maps, the Gibbs sampler and Iterated Conditional Modes (ICM). The method is tested against the ICP considered
as the gold standard for automated rigid registration. Furthermore, the influence of different norms and sampling
point densities is evaluated. The performance of the two methods has been evaluated on data consisting of 178
scanned ear impressions taken from the right ear. To quantify the difference of the two methods we calculate
the registration cost and the mean point to point distance. T-test for common mean are used to determine
the performance of the two methods (supported by a Wilcoxon signed rank test). The performance influence of
sampling density, sampling quantity, and norms is analyzed using a similar method.
Image registration is an important task in most medical imaging applications. Numerous algorithms have been
proposed and some are widely used. However, due to the vast amount of data collected by eg. a computed
tomography (CT) scanner, most registration algorithms are very slow and memory consuming. This is a huge
problem especially in atlas building, where potentially hundreds of registrations are performed. This paper
describes an approach for accelerated image registration. A grid-based warp function proposed by Cootes and
Twining, parameterized by the displacement of the grid-nodes, is used. Using a coarse-to-fine approach, the
composition of small diffeomorphic warps, results in a final diffeomorphic warp. Normally the registration is
done using a standard gradient-based optimizer, but to obtain a fast algorithm the optimization is formulated in
the inverse compositional framework proposed by Baker and Matthews. By switching the roles of the target and
the input volume, the Jacobian and the Hessian can be pre-calculated resulting in a very efficient optimization
algorithm. By exploiting the local nature of the grid-based warp, the storage requirements of the Jacobian and
the Hessian can be minimized. Furthermore, it is shown that additional constraints on the registration, such
as the location of markers, are easily embedded in the optimization. The method is applied on volumes built
from CT-scans of pig-carcasses, and results show a two-fold increase in speed using the inverse compositional
approach versus the traditional gradient-based method.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.