Figure 2 - uploaded by Peter Rogelj
Content may be subject to copyright.
Source publication
We describe the evaluation of a non-rigid image registration method for multi-modal data. The evaluation is made di#cult by the absence of gold standard test data, for which the true transformation from one image to another is known. Di#erent approaches have been used to deal with this deficiency, e.g., by using synthetically warped data, by compar...
Context in source publication
Context 1
... this way the displacements are filtered throughout the iteration process, such that old forces contribute less than later ones. The undesirable side effect of this filtering is that as the external forces go to zero, the image gradually returns back to its undeformed configuration. Thus, additional external forces are needed to sustain the deformed condition. Thus, we simply combine the two spatial deformation models ( Fig. ...
Similar publications
We present novel algorithms for learning structured predictors from instances with multiple labels in the presence of noise. The proposed algorithms improve performance on two standard NLP tasks when we have a small amount of training data (low quantity) and when the labels are noisy (low quality). In these settings, the methods improve performance...
This paper describes new, simple, recursive methods of construction for orientable sequences, i.e. periodic binary sequences in which any n-tuple occurs at most once in a period in either direction. As has been previously described, such sequences have potential applications in automatic position-location systems, where the sequence is encoded onto...
Citations
... Such test data is typically constructed by applying a set of known deformations to a real image. This artificially-deformed data is then registered, and evaluation is based on comparison between the deformation fields recovered by the NRR and those that were originally applied [4], [5]. This approach has two limitations: (i) since the images to be registered are derived from the same source image, they are structurally similar and have similar appearance; (ii) it is difficult to ensure that the deformations used to generate the test images are typical of the anatomical variability found in real data. ...
... Examples of this approach involve measurement of the misregistration of anatomical regions of significance [3], and the overlap between anatomically equivalent regions obtained using segmentation. This process is either manual or semi-automatic [3], [4]. Although these methods cover a general range of applications, they are labour-intensive, and the manual component often suffers from excessive subjectivity. ...
We present a generic method for assessing the quality of non-rigid registration (NRR) algorithms, that does not depend on the existence of any ground truth, but depends solely on the data itself. The data is a set of images. The output of any NRR of such a set of images is a dense correspondence across the whole set. Given such a dense correspondence, it is possible to build various generative statistical models of appearance variation across the set. We show that evaluating the quality of the registration can be mapped to the problem of evaluating the quality of the resultant statistical model. The quality of the model entails a comparison between the model and the image data that was used to construct it. It should be noted that this approach does not depend on the specifics of the registration algorithm used (i.e., whether a groupwise or pairwise algorithm was used to register the set of images), or on the specifics of the modelling approach used. We derive an index of image model specificity that can be used to assess image model quality, and hence the quality of registration. This approach is validated by comparing our assessment of registration quality with that derived from ground truth anatomical labeling. We demonstrate that our approach is capable of assessing NRR reliably without ground truth. Finally, to demonstrate the practicality of our method, different NRR algorithms -- both pairwise and groupwise -- are compared in terms of their performance on 3D MR brain data.
... Early methods for validation focused on visual assessment of the registration results. This was performed by viewing image pairs and comparing their contour overlays, alternate pixel displays, anatomical landmarks, or analytical models (Guehring, 2001;Schnabel et al., 2001;Rogelj et al., 2002;Pappas et al., 2005). These techniques were mostly based on algorithms such as template matching and pattern classification in 2D space. ...
This paper proposes a novel deep learning architecture called ValidNet to automatically validate 3D surface
registration algorithms for object recognition and pose determination tasks. The performance of many tasks
such as object detection mainly depends on the applied registration algorithms, which themselves are susceptible to local minima. Revealing this tendency and verifying the success of registration algorithms is a difficult
task. We treat this as a classification problem, and propose a two-class classifier to distinguish clearly between
true positive and false positive instances. Our proposed ValidNet deploys a shared mlp architecture which
works on the raw and unordered numeric data of scene and model points. This network is able to perform two
fundamental tasks of feature extraction and similarity matching using the powerful capability of deep neural
network. Experiments on a large synthetic datasets show that the proposed method can effectively be used in
automatic validation of registration.
... This repository consists of a simulated brain database providing synthetic MRIs computationally generated. The BrainWeb repository has been widely used by the IR research community [55]. In order to consider different degrees of complexity, some images include noise and multiple sclerosis lesions, as detailed 650 in Table 4. ...
In medical imaging there is a special interest in relating information from different images frequently used for diagnosis or treatment. Image registration (IR) involves the transformation of different sets of image data having a shared content into a common coordinate system. The estimation of the optimal transformation is modelled either as a combinatorial or a numerical optimization problem. Since traditional IR methods are constrained by several limitations, other optimization methods have been recently proposed to overcome such shortcomings. In this contribution, we consider the use of a recently proposed and high performance bio-inspired meta-heuristic: the Coral Reef Optimization Algorithm with Substrate Layers (CRO-SL). We adapt the algorithm to the real-coding IR problem variant following both feature-based and intensity-based designs, and perform two thorough experimental studies. Such studies focus on both mono-modal and inter-modal scenarios where the images suffer different types of 3D affine transformations to validate our proposal. The new proposal is benchmarked with state-of-the-art evolutionary and non-evolutionary IR methods. The results show that CRO-SL is a very competitive approach in terms of its robustness, accuracy, and efficiency.
... The images used in this experiment are the same used in [9], which were obtained from the BrainWeb database at McGill University [11], consisting in simulated brain magnetic resonance images (MRI). The BrainWeb repository is wellknown by the IR research community [30]. In order to consider ...
... Each metric exhibits strengths and weaknesses (33,(37)(38)(39) but none of them have been proven best above the others in terms performance. This suggests that the use of a combination of metrics is worth investigating. ...
The aim of this work is to provide insights into multiple metrics clinical validation of deformable image registration and contour propagation methods in 4D lung radiotherapy planning. The following indices were analyzed and compared: Volume Difference (VD), Dice Similarity Coefficient (DSC), Positive Predictive Value (PPV) and Surface Distances (SD). The analysis was performed on three patient datasets, using as reference a ground-truth volume generated by means of Simultaneous Truth And Performance Level Estimation (STAPLE) algorithm from the outlines of five experts. Significant discrepancies in the quality assessment provided by the different metrics in all the examined cases were found. Metrics sensitivity was more evident in presence of image artifacts and particularly for tubular anatomical structures, such as esophagus or spinal cord. Volume Differences did not account for position and DSC exhibited criticalities due to its intrinsic symmetry (i.e. over- and under-estimation of the reference contours cannot be discriminated) and dependency on the total volume of the structure. PPV analysis showed more robust performance, as each voxel concurs to the classification of the propagation, but was not able to detect inclusion of propagated and ground-truth volumes. Mesh distances could interpret the actual shape of the structures, but might report higher mismatches in case of large local differences in the contour surfaces. According to our study, the combination of VD and SD for the validation of contour propagation algorithms in 4D could provide the necessary failure detection accuracy.
... In the past, evaluating the performance of deformable registration algorithms has been accomplished through use of simulated deformation fields, demonstration of average images in atlas space, and comparison with small numbers of manually selected landmarks [41][42][43][44][45]. With the advent of the Non-Rigid Image Registration Evaluation Project (NIREP) [27,28], it is now possible to additionally evaluate the performance of a deformable registration algorithm through its ability to transfer labels between images. ...
Image registration is a crucial step in many medical image analysis procedures such as image fusion, surgical planning, segmentation and labeling, and shape comparison in population or longitudinal studies. A new approach to volumetric intersubject deformable image registration is presented. The method, called Mjolnir, is an extension of the highly successful method HAMMER. New image features in order to better localize points of correspondence between the two images are introduced as well as a novel approach to generate a dense displacement field based upon the weighted diffusion of automatically derived feature correspondences. An extensive validation of the algorithm was performed on T1-weighted SPGR MR brain images from the NIREP evaluation database. The results were compared with results
generated by HAMMER and are shown to yield significant improvements in cortical alignment as well as
reduced computation time.
... If we cannot expect every brain to have a one-to-one mapping with every other brain, then if possible we need to compare similar brains. This can easily lead to the confound where image correspondence is mistaken for anatomic correspondence (Crum et al., 2003;Rogelj et al., 2002). Choosing a representative brain with which to establish correspondences with a given brain results in a Catch-22 where determining similarities itself entails determining correspondences between the brains. ...
All fields of neuroscience that employ brain imaging need to communicate their results with reference to anatomical regions. In particular, comparative morphometry and group analysis of functional and physiological data require coregistration of brains to establish correspondences across brain structures. It is well established that linear registration of one brain to another is inadequate for aligning brain structures, so numerous algorithms have emerged to nonlinearly register brains to one another. This study is the largest evaluation of nonlinear deformation algorithms applied to brain image registration ever conducted. Fourteen algorithms from laboratories around the world are evaluated using 8 different error measures. More than 45,000 registrations between 80 manually labeled brains were performed by algorithms including: AIR, ANIMAL, ART, Diffeomorphic Demons, FNIRT, IRTK, JRD-fluid, ROMEO, SICLE, SyN, and four different SPM5 algorithms ("SPM2-type" and regular Normalization, Unified Segmentation, and the DARTEL Toolbox). All of these registrations were preceded by linear registration between the same image pairs using FLIRT. One of the most significant findings of this study is that the relative performances of the registration methods under comparison appear to be little affected by the choice of subject population, labeling protocol, and type of overlap measure. This is important because it suggests that the findings are generalizable to new subject populations that are labeled or evaluated using different labeling protocols. Furthermore, we ranked the 14 methods according to three completely independent analyses (permutation tests, one-way ANOVA tests, and indifference-zone ranking) and derived three almost identical top rankings of the methods. ART, SyN, IRTK, and SPM's DARTEL Toolbox gave the best results according to overlap and distance measures, with ART and SyN delivering the most consistently high accuracy across subjects and label sets. Updates will be published on the http://www.mindboggle.info/papers/ website.
... Quantitative evaluation of image registration is a very difficult task. Current methods that are used include use of analytically deformed images some of which consider the biomechanical properties of the patient, [15][16][17][18][19][20] and/or use of contours and bifurcations identified on both image sets for comparison. 15,[21][22][23][24] A voxel-based evaluation method was proposed by Zhong et al. which automatically detects a region in an image where the registration is not performing well using a finite-element-based elastic framework to calculate the unbalanced energy in each voxel after substitution of the displacement vector field. ...
... These values are on the same order as accuracies reported in the literature. [15][16][17][18][19][20][21][22][23][24][25] The maximum error, however, showed a wider range, from 5.1 to 15.4 mm, indicating nonuniformity in the results of deformable image registration and the potential for large regional inaccuracy in alignment in spite of overall acceptable accuracy. Table II ͒ as well as the mean ͑d k ͒, standard deviation ͑ k ͒, and the maximum 3D error ͑d k max ͒ for each registration technique. ...
The looming potential of deformable alignment tools to play an integral role in adaptive radiotherapy suggests a need for objective assessment of these complex algorithms. Previous studies in this area are based on the ability of alignment to reproduce analytically generated deformations applied to sample image data, or use of contours or bifurcations as ground truth for evaluation of alignment accuracy. In this study, a deformable phantom was embedded with 48 small plastic markers, placed in regions varying from high contrast to roughly uniform regional intensity, and small to large regional discontinuities in movement. CT volumes of this phantom were acquired at different deformation states. After manual localization of marker coordinates, images were edited to remove the markers. The resulting image volumes were sent to five collaborating institutions, each of which has developed previously published deformable alignment tools routinely in use. Alignments were done, and applied to the list of reference coordinates at the inhale state. The transformed coordinates were compared to the actual marker locations at exhale. A total of eight alignment techniques were tested from the six institutions. All algorithms performed generally well, as compared to previous publications. Average errors in predicted location ranged from 1.5 to 3.9 mm, depending on technique. No algorithm was uniformly accurate across all regions of the phantom, with maximum errors ranging from 5.1 to 15.4 mm. Larger errors were seen in regions near significant shape changes, as well as areas with uniform contrast but large local motion discontinuity. Although reasonable accuracy was achieved overall, the variation of error in different regions suggests caution in globally accepting the results from deformable alignment.
... In classes 2, 3, and 4, the average distance would equal zero if the images were perfectly registered. (5) Finally, this class quantifies error by computing the overlap of corresponding structures that are defined in the registered images (Rogelj et al.,2002; Jannin et al., 2006; Hellier et al., 2003). The delineated structures overlap completely if the images are registered precisely. ...
Images from different individuals typically cannot be registered precisely because anatomical features within the images differ across the people imaged and because the current methods for image registration have inherent technological limitations that interfere with perfect registration. Quantifying the inevitable error in image registration is therefore of crucial importance in assessing the effects that image misregistration may have on subsequent analyses in an imaging study. We have developed a mathematical framework for quantifying errors in registration by computing the confidence intervals of the estimated parameters (3 translations, 3 rotations, and 1 global scale) for the similarity transformation.
... Linear and non-linear partial derivatives are computed in each iteration, and parameter updates of linear parameters and local displacements are added using a linearly weighted convex combination [22]. Consistency of the non-linear deformation field is ensured by explicit a posteriori regularization, using either a discretized Gaussian [23,24] or linear elastic convolution kernel [25]. Finally, the computed deformation field is applied using trilinear interpolation. ...
Automated full-field 3-D breast ultrasound (3DBUS) has a high potential as a reproducible method for screening and intervention. Consecutive linear transducer scans yield a consistent breast ultrasound volume, yet individual slices are prone to tissue deformation and motion. To compensate resulting image distortions, we propose an efficient non-rigid registration method applied sequentially to pairs of 3DBUS volume slices, optionally either on-line or in post-processing. A quantitative evaluation of the method on synthetic deformations shows subvoxel registration accuracy. First application to clinical breast US images and preliminary results confirmed effectiveness and accuracy of the method.
