Darko Zikic

The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)

In this paper we report the set-up and results of the Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) organized in conjunction with the MICCAI 2012 and 2013 conferences. Twenty state-of-the-art tumor segmentation algorithms were applied to a set of 65 multi-contrast MR scans of low- and high-grade glioma patients - manually annotated by up to four raters - and to 65 comparable scans generated using tumor image simulation software. Quantitative evaluations revealed considerable disagreement between the human raters in segmenting various tumor sub-regions (Dice scores in the range 74%-85%), illustrating the difficulty of this task. We found that different algorithms worked best for different sub-regions (reaching performance comparable to human inter-rater variability), but that no single algorithm ranked in the top for all sub-regions simultaneously. Fusing several good algorithms using a hierarchical majority vote yielded segmentations that consistently ranked above all individual algorithms, indicating remaining opportunities for further methodological improvements. The BRATS image data and manual annotations continue to be publicly available through an online evaluation system as an ongoing benchmarking resource.

Decision forests for tissue-specific segmentation of high-grade gliomas in multi-channel MR.

We present a method for automatic segmentation of high-grade gliomas and their subregions from multi-channel MR images. Besides segmenting the gross tumor, we also differentiate between active cells, necrotic core, and edema. Our discriminative approach is based on decision forests using context-aware spatial features, and integrates a generative model of tissue appearance, by using the probabilities obtained by tissue-specific Gaussian mixture models as additional input for the forest. Our method classifies the individual tissue types simultaneously, which has the potential to simplify the classification task. The approach is computationally efficient and of low model complexity. The validation is performed on a labeled database of 40 multi-channel MR images, including DTI. We assess the effects of using DTI, and varying the amount of training data. Our segmentation results are highly accurate, and compare favorably to the state of the art.

Deformable 2D-3D Registration of Vascular Structures in a One View Scenario

Alignment of angiographic 3D scans to 2D projections is an important issue for 3D depth perception and navigation during interventions. Currently, in a setting where only one 2D projection is available, methods employing a rigid transformation model present the state of the art for this problem. In this work, we introduce a method capable of deformably registering 3D vessel structures to a respective single projection of the scene. Our approach addresses the inherent ill-posedness of the problem by incorporating a priori knowledge about the vessel structures into the formulation. We minimize the distance between the 2D points and corresponding projected 3D points together with regularization terms encoding the properties of length preservation of vessel structures and smoothness of deformation. We demonstrate the performance and accuracy of the proposed method by quantitative tests on synthetic examples as well as real angiographic scenes.

Vertebrae Localization in Pathological Spine CT via Dense Classification from Sparse Annotations

Accurate localization and identification of vertebrae in spinal imaging is crucial for the clinical tasks of diagnosis, surgical planning, and post-operative assessment. The main difficulties for automatic methods arise from the frequent presence of abnormal spine curvature, small field of view, and image artifacts caused by surgical implants. Many previous methods rely on parametric models of appearance and shape whose performance can substantially degrade for pathological cases. We propose a robust localization and identification algorithm which builds upon supervised classification forests and avoids an explicit parametric model of appearance. We overcome the tedious requirement for dense annotations by a semi-automatic labeling strategy. Sparse centroid annotations are transformed into dense probabilistic labels which capture the inherent identification uncertainty. Using the dense labels, we learn a discriminative centroid classifier based on local and contextual intensity features which is robust to typical characteristics of spinal pathologies and image artifacts. Extensive evaluation is performed on a challenging dataset of 224 spine CT scans of patients with varying pathologies including high-grade scoliosis, kyphosis, and presence of surgical implants. Additionally, we test our method on a heterogeneous dataset of another 200, mostly abdominal, CTs. Quantitative evaluation is carried out with respect to localization errors and identification rates, and compared to a recently proposed method. Our approach is efficient and outperforms state-of-the-art on pathological cases.

Encoding atlases by randomized classification forests for efficient multi-atlas label propagation

We propose a method for multi-atlas label propagation (MALP) based on encoding the individual atlases by randomized classification forests. Most current approaches perform a non-linear registration between all atlases and the target image, followed by a sophisticated fusion scheme. While these approaches can achieve high accuracy, in general they do so at high computational cost. This might negatively affect the scalability to large databases and experimentation. To tackle this issue, we propose to use a small and deep classification forest to encode each atlas individually in reference to an aligned probabilistic atlas, resulting in an Atlas Forest (AF). Our classifier-based encoding differs from current MALP approaches, which represent each point in the atlas either directly as a single image/label value pair, or by a set of corresponding patches. At test time, each AF produces one probabilistic label estimate, and their fusion is done by averaging. Our scheme performs only one registration per target image, achieves good results with a simple fusion scheme, and allows for efficient experimentation. In contrast to standard forest schemes, in which each tree would be trained on all atlases, our approach retains the advantages of the standard MALP framework. The target-specific selection of atlases remains possible, and incorporation of new scans is straightforward without retraining. The evaluation on four different databases shows accuracy within the range of the state of the art at a significantly lower running time.

Atlas Encoding by Randomized Forests for Efficient Label Propagation

We propose a method for multi-atlas label propagation based on encoding the individual atlases by randomized classification forests. Most current approaches perform a non-linear registration between all atlases and the target image, followed by a sophisticated fusion scheme. While these approaches can achieve high accuracy, in general they do so at high computational cost. This negatively affects the scalability to large databases and experimentation. To tackle this issue, we propose to use a small and deep classification forest to encode each atlas individually in reference to an aligned probabilistic atlas, resulting in an Atlas Forest (AF). At test time, each AF yields a probabilistic label estimate, and fusion is done by averaging. Our scheme performs only one registration per target image, achieves good results with a simple fusion scheme, and allows for efficient experimentation. In contrast to standard forest schemes, incorporation of new scans is possible without retraining, and target-specific selection of atlases remains possible. The evaluation on three different databases shows accuracy at the level of the state of the art, at a significantly lower runtime.

Is Synthesizing MRI Contrast Useful for Inter-modality Analysis?

Availability of multi-modal magnetic resonance imaging (MRI) databases opens up the opportunity to synthesize different MRI contrasts without actually acquiring the images. In theory such synthetic images have the potential to reduce the amount of acquisitions to perform certain analyses. However, to what extent they can substitute real acquisitions in the respective analyses is an open question. In this study, we used a synthesis method based on patch matching to test whether synthetic images can be useful in segmentation and inter-modality cross-subject registration of brain MRI. Thirty-nine T1 scans with 36 manually labeled structures of interest were used in the registration and segmentation of eight proton density (PD) scans, for which ground truth T1 data were also available. The results show that synthesized T1 contrast can considerably enhance the quality of non-linear registration compared with using the original PD data, and it is only marginally worse than using the original T1 scans. In segmentation, the relative improvement with respect to using the PD is smaller, but still statistically significant.

Linear intensity-based image registration by Markov random fields and discrete optimization

We propose a framework for intensity-based registration of images by linear transformations, based on a discrete Markov random field (MRF) formulation. Here, the challenge arises from the fact that optimizing the energy associated with this problem requires a high-order MRF model. Currently, methods for optimizing such high-order models are less general, easy to use, and efficient, than methods for the popular second-order models. Therefore, we propose an approximation to the original energy by an MRF with tractable second-order terms. The approximation at a certain point p in the parameter space is the normalized sum of evaluations of the original energy at projections of p to two-dimensional subspaces. We demonstrate the quality of the proposed approximation by computing the correlation with the original energy, and show that registration can be performed by discrete optimization of the approximated energy in an iteration loop. A search space refinement strategy is employed over iterations to achieve sub-pixel accuracy, while keeping the number of labels small for efficiency. The proposed framework can encode any similarity measure is robust to the settings of the internal parameters, and allows an intuitive control of the parameter ranges. We demonstrate the applicability of the framework by intensity-based registration, and 2D-3D registration of medical images. The evaluation is performed by random studies and real registration tasks. The tests indicate increased robustness and precision compared to corresponding standard optimization of the original energy, and demonstrate robustness to noise. Finally, the proposed framework allows the transfer of advances in MRF optimization to linear registration problems.

Modality Propagation: Coherent Synthesis of Subject-Specific Scans with Data-Driven Regularization

We propose a general database-driven framework for coherent synthesis of subject-specific scans of desired modality, which adopts and generalizes the patch-based label propagation (LP) strategy. While modality synthesis has received increased attention lately, current methods are mainly tailored to specific applications. On the other hand, the LP framework has been extremely successful for certain segmentation tasks, however, so far it has not been used for estimation of entities other than categorical segmentation labels. We approach the synthesis task as a modality propagation, and demonstrate that with certain modifications the LP framework can be generalized to continuous settings providing coherent synthesis of different modalities, beyond segmentation labels. To achieve high-quality estimates we introduce a new data-driven regularization scheme, in which we integrate intermediate estimates within an iterative search-and-synthesis strategy. To efficiently leverage population data and ensure coherent synthesis, we employ a spatio-population search space restriction. In experiments, we demonstrate the quality of synthesis of different MRI signals (T2 and DTI-FA) from a T1 input, and show a novel application of modality synthesis for abnormality detection in multi-channel MRI of brain tumor patients.

Fast deformable registration of 3d-ultrasound data using a variational approach

We present an intensity based deformable registration algorithm for 3D ultrasound data. The proposed method uses a variational approach and combines the characteristics of a multilevel algorithm and the properties of ultrasound data in order to provide a fast and accurate deformable registration method. In contrast to previously proposed approaches, we use no feature points and no interpolation technique, but compute a dense displacement field directly. We demonstrate that this approach, although it includes solving large PDE systems, reduces the computation time if implemented using efficient numerical techniques. The performance of the algorithm is tested on multiple 3D US images of the liver. Validation is performed by simulations, similarity comparisons between original and deformed images, visual inspection of the displacement fields and visual assessment of the deformed images by physicians.