Localized Statistical Shape Models for Large-Scale Problems With Few Training Data.

OBJECTIVE: Statistical shape models have been successfully used in numerous biomedical image analysis applications where prior shape information is helpful such as organ segmentation or data augmentation when training deep learning models. However, training such models requires large data sets, which are often not available and, hence, shape models frequently fail to represent local details of unseen shapes. This work introduces a kernel-based method to alleviate this problem via so-called model localization. It is specifically designed to be used in large-scale shape modeling scenarios like deep learning data augmentation and fits seamlessly into the classical shape modeling framework. METHOD: Relying on recent advances in multi-level shape model localization via distance-based covariance matrix manipulations and Grassmannian-based level fusion, this work proposes a novel and computationally efficient kernel-based localization technique. Moreover, a novel way to improve the specificity of such models via normalizing flow-based density estimation is presented. RESULTS: The method is evaluated on the publicly available JSRT/SCR chest X-ray and IXI brain data sets. The results confirm the effectiveness of the kernelized formulation and also highlight the models' improved specificity when utilizing the proposed density estimation method. CONCLUSION: This work shows that flexible and specific shape models from few training samples can be generated in a computationally efficient way by combining ideas from kernel theory and normalizing flows. SIGNIFICANCE: The proposed method together with its publicly available implementation allows to build shape models from few training samples directly usable for applications like data augmentation.

A systematic comparison of generative models for medical images.

PURPOSE: This work aims for a systematic comparison of popular shape and appearance models. Here, two statistical and four deep-learning-based shape and appearance models are compared and evaluated in terms of their expressiveness described by their generalization ability and specificity as well as further properties like input data format, interpretability and latent space distribution and dimension. METHODS: Classical shape models and their locality-based extension are considered next to autoencoders, variational autoencoders, diffeomorphic autoencoders and generative adversarial networks. The approaches are evaluated in terms of generalization ability, specificity and likeness depending on the amount of training data. Furthermore, various latent space metrics are presented in order to capture further major characteristics of the models. RESULTS: The experimental setup showed that locality statistical shape models yield best results in terms of generalization ability for 2D and 3D shape modeling. However, the deep learning approaches show strongly improved specificity. In the case of simultaneous shape and appearance modeling, the neural networks are able to generate more realistic and diverse appearances. A major drawback of the deep-learning models is, however, their impaired interpretability and ambiguity of the latent space. CONCLUSIONS: It can be concluded that for applications not requiring particularly good specificity, shape modeling can be reliably established with locality-based statistical shape models, especially when it comes to 3D shapes. However, deep learning approaches are more worthwhile in terms of appearance modeling.

Image registration and appearance adaptation in non-correspondent image regions for new MS lesions detection.

Manual detection of newly formed lesions in multiple sclerosis is an important but tedious and difficult task. Several approaches for automating the detection of new lesions have recently been proposed, but they tend to either overestimate the actual amount of new lesions or to miss many lesions. In this paper, an image registration convolutional neural network (CNN) that adapts the baseline image to the follow-up image by spatial deformations and simulation of new lesions is proposed. Simultaneously, segmentations of new lesions are generated, which are shown to reliably estimate the real new lesion load and to separate stable and progressive patients. Several applications of the proposed network emerge: image registration, detection and segmentation of new lesions, and modeling of new MS lesions. The modeled lesions offer the possibility to investigate the intensity profile of new lesions.

Deep learning-based simultaneous registration and unsupervised non-correspondence segmentation of medical images with pathologies.

PURPOSE: The registration of medical images often suffers from missing correspondences due to inter-patient variations, pathologies and their progression leading to implausible deformations that cause misregistrations and might eliminate valuable information. Detecting non-corresponding regions simultaneously with the registration process helps generating better deformations and has been investigated thoroughly with classical iterative frameworks but rarely with deep learning-based methods. METHODS: We present the joint non-correspondence segmentation and image registration network (NCR-Net), a convolutional neural network (CNN) trained on a Mumford-Shah-like functional, transferring the classical approach to the field of deep learning. NCR-Net consists of one encoding and two decoding parts allowing the network to simultaneously generate diffeomorphic deformations and segment non-correspondences. The loss function is composed of a masked image distance measure and regularization of deformation field and segmentation output. Additionally, anatomical labels are used for weak supervision of the registration task. No manual segmentations of non-correspondences are required. RESULTS: The proposed network is evaluated on the publicly available LPBA40 dataset with artificially added stroke lesions and a longitudinal optical coherence tomography (OCT) dataset of patients with age-related macular degeneration. The LPBA40 data are used to quantitatively assess the segmentation performance of the network, and it is shown qualitatively that NCR-Net can be used for the unsupervised segmentation of pathologies in OCT images. Furthermore, NCR-Net is compared to a registration-only network and state-of-the-art registration algorithms showing that NCR-Net achieves competitive performance and superior robustness to non-correspondences. CONCLUSION: NCR-Net, a CNN for simultaneous image registration and unsupervised non-correspondence segmentation, is presented. Experimental results show the network's ability to segment non-correspondence regions in an unsupervised manner and its robust registration performance even in the presence of large pathologies.

Memory-efficient GAN-based domain translation of high resolution 3D medical images.

Generative adversarial networks (GANs) are currently rarely applied on 3D medical images of large size, due to their immense computational demand. The present work proposes a multi-scale patch-based GAN approach for establishing unpaired domain translation by generating 3D medical image volumes of high resolution in a memory-efficient way. The key idea to enable memory-efficient image generation is to first generate a low-resolution version of the image followed by the generation of patches of constant sizes but successively growing resolutions. To avoid patch artifacts and incorporate global information, the patch generation is conditioned on patches from previous resolution scales. Those multi-scale GANs are trained to generate realistically looking images from image sketches in order to perform an unpaired domain translation. This allows to preserve the topology of the test data and generate the appearance of the training domain data. The evaluation of the domain translation scenarios is performed on brain MRIs of size 155 × 240 × 240 and thorax CTs of size up to 5123. Compared to common patch-based approaches, the multi-resolution scheme enables better image quality and prevents patch artifacts. Also, it ensures constant GPU memory demand independent from the image size, allowing for the generation of arbitrarily large images.

High-resolution T2-FLAIR and non-contrast CT brain atlas of the elderly.

Normative brain atlases are a standard tool for neuroscience research and are, for example, used for spatial normalization of image datasets prior to voxel-based analyses of brain morphology and function. Although many different atlases are publicly available, they are usually biased with respect to an imaging modality and the age distribution. Both effects are well known to negatively impact the accuracy and reliability of the spatial normalization process using non-linear image registration methods. An important and very active neuroscience area that lacks appropriate atlases is lesion-related research in elderly populations (e.g. stroke, multiple sclerosis) for which FLAIR MRI and non-contrast CT are often the clinical imaging modalities of choice. To overcome the lack of atlases for these tasks and modalities, this paper presents high-resolution, age-specific FLAIR and non-contrast CT atlases of the elderly generated using clinical images.

Patient-specific finite element modeling of respiratory lung motion using 4D CT image data.

Development and optimization of methods for adequately accounting for respiratory motion in radiation therapy of thoracic tumors require detailed knowledge of respiratory dynamics and its impact on corresponding dose distributions. Thus, computer aided modeling and simulation of respiratory motion have become increasingly important. In this article a biophysical approach for modeling respiratory lung motion is described: Major aspects of the process of lung ventilation are formulated as a contact problem of elasticity theory which is solved by finite element methods; lung tissue is assumed to be isotropic, homogeneous, and linearly elastic. A main focus of the article is to assess the impact of biomechanical parameters (values of elastic constants) on the modeling process and to evaluate modeling accuracy. Patient-specific models are generated based on 4D CT data of 12 lung tumor patients. Simulated motion patterns of inner lung landmarks are compared with corresponding motion patterns observed in the 4D CT data. Mean absolute differences between model-based predicted landmark motion and corresponding breathing-induced landmark displacements as observed in the CT data sets are in the order of 3 mm (end expiration to end inspiration) and 2 mm (end expiration to midrespiration). Modeling accuracy decreases with increasing tumor size both locally (landmarks close to tumor) and globally (landmarks in other parts of the lung). The impact of the values of the elastic constants appears to be small. Outcomes show that the modeling approach is an adequate strategy in predicting lung dynamics due to lung ventilation. Nevertheless, the decreased prediction quality in cases of large tumors demands further study of the influence of lung tumors on global and local lung elasticity properties.

Reproducible extraction of local and global parameters for functional analysis of the left ventricle in 4D MR image data.

OBJECTIVES: Left ventricle (LV) segmentation is required to quantify LV volume and mass parameters. Therefore, spatiotemporal Cine MR sequences in the short and long axis of the heart are acquired. Generally, LV segmentation methods consider short-axis sequences only. The reduced resolution in long-axis direction is one of the main reasons for inaccurate parameter extraction in the apical and basal area. The segmentation approach presented combines short- and long-axis information as well as motion tracking to enable the functional LV analysis in 4D MR Image Data. METHODS: First, anatomical landmarks like the mitral valve and the apex are defined in long-axis views in diastolic and systolic phase in order to specify the upper and lower boundary of the LV. Second, motion field approximation using non-linear registration enables the automatic contour propagation to all time points. Third, intersection planes are defined parallel to the mitral valve plane covering the whole ventricle. Finally, the 4D LV surface model is generated appending all in-plane contours. The segmentation results in short-axis images are checked and adjusted interactively and quantitative parameters are extracted. RESULTS: For evaluation the contours of 19 different datasets were traced by two medical experts using a contour drawing tool and the new segmentation tool. The results were compared to evaluate automatic contour propagation, robustness of the segmentation as well as interaction time. CONCLUSION: The automatic contour propagation enables the fast and reproducible generation of a 4D model for the functional analysis of the heart. The interaction time is decreased from approx. 60 minutes to 10 minutes per case. Inter- and intraobserver differences of the extracted parameters are decreased significantly.

Unsupervised pathology detection in medical images using conditional variational autoencoders.

PURPOSE: Pathology detection in medical image data is an important but a rather complicated task. In particular, the big variability of the pathologies is a challenge to automatic detection methods and even to machine learning methods. Supervised algorithms would usually learn the appearance of a single pathological structure based on a large annotated dataset. As such data is not usually available, especially in large amounts, in this work we pursue a different unsupervised approach. METHODS: Our method is based on learning the entire variability of healthy data and detect pathologies by their differences to the learned norm. For this purpose, we use conditional variational autoencoders which learn the reconstruction and encoding distribution of healthy images and also have the ability to integrate certain prior knowledge about the data (condition). RESULTS: Our experiments on different 2D and 3D datasets show that the approach is suitable for the detection of pathologies and deliver reasonable Dice coefficients and AUCs. Also this method can estimate missing correspondences in pathological images and thus can be used as a pre-step to a registration method. Our experiments show improving registration results on pathological data when using this approach. CONCLUSIONS: Overall the presented approach is suitable for a rough pathology detection in medical images and can be successfully used as a preprocessing step to other image processing methods.

Statistical shape modeling of the left ventricle: myocardial infarct classification challenge.

Statistical shape modeling is a powerful tool for visualizing and quantifying geometric and functional patterns of the heart. After myocardial infarction (MI), the left ventricle typically remodels in response to physiological challenges. Several methods have been proposed in the literature to describe statistical shape changes. Which method best characterizes left ventricular remodeling after MI is an open research question. A better descriptor of remodeling is expected to provide a more accurate evaluation of disease status in MI patients. We therefore designed a challenge to test shape characterization in MI given a set of three-dimensional left ventricular surface points. The training set comprised 100 MI patients, and 100 asymptomatic volunteers (AV). The challenge was initiated in 2015 at the Statistical Atlases and Computational Models of the Heart workshop, in conjunction with the MICCAI conference. The training set with labels was provided to participants, who were asked to submit the likelihood of MI from a different (validation) set of 200 cases (100 AV and 100 MI). Sensitivity, specificity, accuracy and area under the receiver operating characteristic curve were used as the outcome measures. The goals of this challenge were to (1) establish a common dataset for evaluating statistical shape modeling algorithms in MI, and (2) test whether statistical shape modeling provides additional information characterizing MI patients over standard clinical measures. Eleven groups with a wide variety of classification and feature extraction approaches participated in this challenge. All methods achieved excellent classification results with accuracy ranges from 0.83 to 0.98. The areas under the receiver operating characteristic curves were all above 0.90. Four methods showed significantly higher performance than standard clinical measures. The dataset and software for evaluation are available from the Cardiac Atlas Project website1.

An optical flow based method for improved reconstruction of 4D CT data sets acquired during free breathing.

Respiratory motion degrades anatomic position reproducibility and leads to issues affecting image acquisition, treatment planning, and radiation delivery. Four-dimensional (4D) computer tomography (CT) image acquisition can be used to measure the impact of organ motion and to explicitly account for respiratory motion during treatment planning and radiation delivery. Modern CT scanners can only scan a limited region of the body simultaneously and patients have to be scanned in segments consisting of multiple slices. A respiratory signal (spirometer signal or surface tracking) is used to reconstruct a 4D data set by sorting the CT scans according to the couch position and signal coherence with predefined respiratory phases. But artifacts can occur if there are no acquired data segments for exactly the same respiratory state for all couch positions. These artifacts are caused by device-dependent limitations of gantry rotation, image reconstruction times and by the variability of the patient's respiratory pattern. In this paper an optical flow based method for improved reconstruction of 4D CT data sets from multislice CT scans is presented. The optical flow between scans at neighboring respiratory states is estimated by a non-linear registration method. The calculated velocity field is then used to reconstruct a 4D CT data set by interpolating data at exactly the predefined respiratory phase. Our reconstruction method is compared with the usually used reconstruction based on amplitude sorting. The procedures described were applied to reconstruct 4D CT data sets for four cancer patients and a qualitative and quantitative evaluation of the optical flow based reconstruction method was performed. Evaluation results show a relevant reduction of reconstruction artifacts by our technique. The reconstructed 4D data sets were used to quantify organ displacements and to visualize the abdominothoracic organ motion.

Live-wire-based segmentation using similarities between corresponding image structures.

A live-wire-based segmentation method that exploits similarities of corresponding object contours is presented. The method accelerates the segmentation process transferring anchor points of segmented reference contours to unsegmented target slices automatically. The target contours are created using the live-wire algorithm trained by features of the reference contours. An automatic contour correction improves the segmentation. Only few user interactions with an intuitive contour editor are necessary. The evaluation using intra- and interpatient transfer shows that 51-73% of interaction time can be saved compared to normal live-wire preserving the segmentation quality.

Multi-resolution multi-object statistical shape models based on the locality assumption.

Statistical shape models learned from a population of previously observed training shapes are nowadays widely used in medical image analysis to aid segmentation or classification. However, providing an appropriate and representative training population of preferably manual segmentations is typically either very labor-intensive or even impossible. Therefore, statistical shape models in practice frequently suffer from the high-dimension-low-sample-size (HDLSS) problem resulting in models with insufficient expressiveness. In this paper, a novel approach for learning representative multi-resolution multi-object statistical shape models from a small number of training samples that adequately model the variability of each individual object as well as their interrelations is presented. The method is based on the assumption of locality, which means that local shape variations have limited effects in distant areas and, therefore, can be modeled independently. This locality assumption is integrated into the standard statistical shape modeling framework by manipulating the sample covariance matrix (non-zero covariances between distant landmarks are set to zero). To allow for multi-object modeling, a method for computing distances between points located on different object shapes is proposed. Furthermore, different levels of locality are introduced by deriving a multi-resolution scheme, which is equipped with a method to combine variability information modeled at different levels into a single shape model. This combined representation of global and local variability in a single shape model allows the use of the classical active shape model strategy for model-based image segmentation. An extensive evaluation based on a public data base of 247 chest radiographs is performed to show the modeling and segmentation capabilities of the proposed approach in single- and multi-object HDLSS scenarios. The new approach is not only compared to the classical shape modeling method but also to three state-of-the-art shape modeling approaches specifically designed to cope with the HDLSS problem. The results show that the new approach significantly outperforms all other approaches in terms of generalization ability and model-based segmentation accuracy.

Statistical appearance models based on probabilistic correspondences.

Model-based image analysis is indispensable in medical image processing. One key aspect of building statistical shape and appearance models is the determination of one-to-one correspondences in the training data set. At the same time, the identification of these correspondences is the most challenging part of such methods. In our earlier work, we developed an alternative method using correspondence probabilities instead of exact one-to-one correspondences for a statistical shape model (Hufnagel et al., 2008). In this work, a new approach for statistical appearance models without one-to-one correspondences is proposed. A sparse image representation is used to build a model that combines point position and appearance information at the same time. Probabilistic correspondences between the derived multi-dimensional feature vectors are used to omit the need for extensive preprocessing of finding landmarks and correspondences as well as to reduce the dependence of the generated model on the landmark positions. Model generation and model fitting can now be expressed by optimizing a single global criterion derived from a maximum a-posteriori (MAP) approach with respect to model parameters that directly affect both shape and appearance of the considered objects inside the images. The proposed approach describes statistical appearance modeling in a concise and flexible mathematical framework. Besides eliminating the demand for costly correspondence determination, the method allows for additional constraints as topological regularity in the modeling process. In the evaluation the model was applied for segmentation and landmark identification in hand X-ray images. The results demonstrate the feasibility of the model to detect hand contours as well as the positions of the joints between finger bones for unseen test images. Further, we evaluated the model on brain data of stroke patients to show the ability of the proposed model to handle partially corrupted data and to demonstrate a possible employment of the correspondence probabilities to indicate these corrupted/pathological areas.

Subpopulation-based correspondence modelling for improved respiratory motion estimation in the presence of inter-fraction motion variations.

Correspondence modelling between low-dimensional breathing signals and internal organ motion is a prerequisite for application of advanced techniques in radiotherapy of moving targets. Patient-specific correspondence models can, for example, be built prior to treatment based on a planning 4D CT and simultaneously acquired breathing signals. Reliability of pre-treatment-built models depends, however, on the degree of patient-specific inter-fraction motion variations. This study investigates whether motion estimation accuracy in the presence of inter-fraction motion variations can be improved using correspondence models that incorporate motion information from different patients. The underlying assumption is that inter-patient motion variations resemble patient-specific inter-fraction motion variations for subpopulations of patients with similar breathing characteristics. The hypothesis is tested by integrating a sparse manifold clustering approach into a regression-based correspondence modelling framework that allows for automated identification of patient subpopulations. The evaluation is based on a total of 73 lung 4D CT data sets, including two cohorts of patients with repeat 4D CT scans (cohort 1: 14 patients; cohort 2: ten patients). The results are consistent for both cohorts: The subpopulation-based modelling approach outperforms general population modelling (models built on all data sets available) as well as pre-treatment-built models trained on only the patient-specific motion information. The results thereby support the hypothesis and illustrate the potential of subpopulation-based correspondence modelling.

Generation of 4D CT image data and analysis of lung tumour mobility during the breathing cycle.

The mobility of lung tumours during the breathing cycle is a source of error in radiotherapy treatment planning. Spatio-temporal CT data sets can be used to measure the movement of lung tumours caused by breathing. Because modern CT scanners can only scan a limited region of the body simultaneously at different times, patients have to be scanned in segments consisting of multiple slices. For studying free breathing motion multislice CT scans can be collected simultaneously with digital spirometry over several breathing cycles. The image data set is assembled by sorting the free breathing multislice CT scans according to the couch position and the tidal volume. But artefacts can occur because there are no data segments for exactly the same tidal volume and all couch positions. In this paper, a non-linear registration method is used to interpolate and reconstruct 4D CT data sets from multislice CT scans in high quality. The non-linear registration estimates a velocity field between successive scans, which is used to reconstruct a 4D CT data set by interpolating data at user-defined tidal volumes. By this technique, artefacts can be reduced significantly. Furthermore, the reconstructed 4D CT data sets are used for studying the motion of lung tumours during the respiratory cycle. The reconstructed 4D data sets of 4 patients were used to quantify the individual lung tumour motion as well as to estimate the tumour's appearance probability during a breathing cycle.

Simulation of mammographic breast compression in 3D MR images using ICP-based B-spline deformation for multimodality breast cancer diagnosis.

PURPOSE: Multimodality mammography using conventional 2D mammography and dynamic contrast-enhanced 3D magnetic resonance imaging (DCE-MRI) is frequently performed for breast cancer detection and diagnosis. Combination of both imaging modalities requires superimposition of corresponding structures in mammograms and MR images. This task is challenging due to large differences in (1) dimensionality and spatial resolution, (2) variations in tissue contrast, as well as (3) differences in breast orientation and deformation during the image acquisition. A new method for multimodality breast image registration was developed and tested. METHODS: Combined diagnosis of mammograms and MRI datasets was achieved by simulation of mammographic breast compression to overcome large differences in breast deformation. Surface information was extracted from the 3D MR image, and back-projection of the 2D breast contour in the mammogram was done. B-spline-based 3D/3D surface-based registration was then used to approximate mammographic breast compression. This breast deformation simulation was performed on 14 MRI datasets with 19 corresponding mammograms. The results were evaluated by comparison with distances between corresponding structures identified by an expert observer. RESULTS: The evaluation revealed an average distance of 6.46 mm between corresponding structures, when an optimized initial alignment between both image datasets is performed. Without the optimization, the accuracy is 9.12 mm. CONCLUSION: A new surface-based method that approximates the mammographic deformation due to breast compression without using a specific complex model needed for finite-element-based methods was developed and tested with favorable results. The simulated compression can serve as foundation for a point-to-line correspondence between 2D mammograms and 3D MR image data.

Simulation of spatiotemporal CT data sets using a 4D MRI-based lung motion model.

PURPOSE: Four-dimensional CT imaging is widely used to account for motion-related effects during radiotherapy planning of lung cancer patients. However, 4D CT often contains motion artifacts, cannot be used to measure motion variability, and leads to higher dose exposure. In this article, we propose using 4D MRI to acquire motion information for the radiotherapy planning process. From the 4D MRI images, we derive a time-continuous model of the average patient-specific respiratory motion, which is then applied to simulate 4D CT data based on a static 3D CT. METHODS: The idea of the motion model is to represent the average lung motion over a respiratory cycle by cyclic B-spline curves. The model generation consists of motion field estimation in the 4D MRI data by nonlinear registration, assigning respiratory phases to the motion fields, and applying a B-spline approximation on a voxel-by-voxel basis to describe the average voxel motion over a breathing cycle. To simulate a patient-specific 4D CT based on a static CT of the patient, a multi-modal registration strategy is introduced to transfer the motion model from MRI to the static CT coordinates. RESULTS: Differences between model-based estimated and measured motion vectors are on average 1.39 mm for amplitude-based binning of the 4D MRI data of three patients. In addition, the MRI-to-CT registration strategy is shown to be suitable for the model transformation. CONCLUSIONS: The application of our 4D MRI-based motion model for simulating 4D CT images provides advantages over standard 4D CT (less motion artifacts, radiation-free). This makes it interesting for radiotherapy planning.

Estimation of lung motion fields in 4D CT data by variational non-linear intensity-based registration: A comparison and evaluation study.

Accurate and robust estimation of motion fields in respiration-correlated CT (4D CT) images, usually performed by non-linear registration of the temporal CT frames, is a precondition for the analysis of patient-specific breathing dynamics and subsequent image-supported diagnostics and treatment planning. In this work, we present a comprehensive comparison and evaluation study of non-linear registration variants applied to the task of lung motion estimation in thoracic 4D CT data. In contrast to existing multi-institutional comparison studies (e.g. MIDRAS and EMPIRE10), we focus on the specific but common class of variational intensity-based non-parametric registration and analyze the impact of the different main building blocks of the underlying optimization problem: the distance measure to be minimized, the regularization approach and the transformation space considered during optimization. In total, 90 different combinations of building block instances are compared. Evaluated on proprietary and publicly accessible 4D CT images, landmark-based registration errors (TRE) between 1.14 and 1.20 mm for the most accurate registration variants demonstrate competitive performance of the applied general registration framework compared to other state-of-the-art approaches for lung CT registration. Although some specific trends can be observed, effects of interchanging individual instances of the building blocks on the TRE are in general rather small (no single outstanding registration variant existing); the same level of accuracy is, however, associated with significantly different degrees of motion field smoothness and computational demands. Consequently, the building block combination of choice will depend on application-specific requirements on motion field characteristics.

3D cerebrovascular segmentation combining fuzzy vessel enhancement and level-sets with anisotropic energy weights.

The aim of this work is to present and evaluate a level-set segmentation approach with vesselness-dependent anisotropic energy weights, which focuses on the exact segmentation of malformed as well as small vessels from time-of-flight (TOF) magnetic resonance angiography (MRA) datasets. In a first step, a vesselness filter is used to calculate the vesselness dataset, which quantifies the likeliness of each voxel to belong to a bright tubular-shaped structure and estimate the corresponding vessel directions from a given TOF dataset. The vesselness and TOF datasets are then combined using fuzzy-logic and used for initialization of a variational level-set method. The proposed level-set model has been extended in a way that the weight of the internal energy is locally adapted based on the vessel direction information. Here, the main idea is to weight the internal energy lower if the gradient direction of the level-set is similar to the direction of the eigenvector extracted by the vesselness filter. Furthermore, an additional vesselness force has been integrated in the level-set formulation. The proposed method was evaluated based on ten TOF MRA datasets from patients with an arteriovenous malformation. Manual segmentations from two observers were available for each dataset and used for quantitative comparison. The evaluation revealed that the proposed method yields significantly better segmentation results than four other state-of-the-art segmentation methods tested. Furthermore, the segmentation results are within the range of the inter-observer variation. In conclusion, the proposed method allows an improved delineation of small vessels, especially of those represented by low intensities and high surface curvatures.