Bildverarbeitung für die Medizin 2017

Big data are dramatically increasing the possibilities for prevention, cure and care, and changing the landscape of the healthcare system. Will artificial intelligence make doctors obsolete or give them more possibilities? Will citizens be delivered into the hands of anonymous information systems or will they gain more control over their personal health. It is difficult to predict the speed of change and impact of both big data and artificial intelligence on health care, but it is clear that changes will be tremendous.

Despite a tremendous advance in knowledge in informatics and specifically in medical image analysis in the last decades, the mainstay of current imagingbased diagnostics remains an artisan hand-work performed by medical doctors. Clearly, a translational gap is present between basic sciences and diagnostic clinical routine that hampers the process of industrialization and automation that has pervasively changed the face of many other jobs, blue and white collar alike.

In the last years, deep convolutional networks have outperformed the state of the art in many visual recognition tasks. A central challenge for its wide adoption in the bio-medical imaging field is the limited amount of annotated training images. In this talk, I will present our u-net for biomedical image segmentation. The architecture consists of an analysis path and a synthesis path with additional shortcut-connections.

Neural networks have recently taken the field of machine learning by storm. Their success rests upon the availability of high performance computing hardware which allows to train very wide and deep networks. Traditional neural networks have very limited biological realism. Recent work on more brain-like hardware architectures has led to first large-scale implementations of neuromorphic computing systems. In the keynote, guiding design principles of neuromorphic machines, their application and performance as well as future plans are discussed.

The increasing networking of data systems in medicine leads not only to a modern interdisciplinarity in the sense of a cooperation of different medical disciplines, but also poses new challenges to the concept of the operating room as building infrastructure. The surgical operating room of the future augments its reality, away from pure space characteristics, to an intelligent and communicative interactive platform. In analogy to a smart phone which becomes more a platform for different application beyond the function as a pure telephone device, the operating room of the future serves as sensor and actuator at the same time. A modular architecture with open interfaces for image acquisition and analysis, interaction and visualization, supports the integration and combination of heterogeneous data that are generated in a hospital and operating room environment.

Deep Learning (DL) represents a key technological innovation in the field of machine learning. Recent advancements have attracted much attention by showing substantial improvements in a wide range of applications such as image recognition, speech recognition, natural language processing and artificial intelligence. In some cases the performance even surpasses human accuracy, which motivated the introduction of a series of DL-based software products and automatization solutions (for example Apple Siri, Google Now, Google Autonomous Driving etc.). The same success also echoes in the research efforts of the medical imaging community. However, in this case several constraints such as data-availability, inherent data noise or lack of labeled data directly affect the pace of advancements.

In der Minimal Invasiven Chirurgie (MIC), auch Schlüssellochchirurgie oder laparoskopische Chirurgie genannt, wird im Gegensatz zur offenen Chirurgie nur über kleine Schnitte operiert. über spezielle Zugänge durch die Bauchdecke werden eine Kamera und lange stabförmige Instrumente in den Bauchraum eingef ührt. Das Verfahren ist durch den Verzicht auf große Schnittwunden besonders schonend für den Patienten und führt häufig zur schnelleren Erholung von der Operation. Die laparoskopischen Operationstechniken haben sich in den letzten Jahren in immer mehr Bereichen durchgesetzt und gewinnen zunehmend an Bedeutung, Verbreitung und Beliebtheit. Für manche Operationen ist die laparoskopische Technik inzwischen schon der etablierte Standard, wie z.B. für die Gallenblasenentfernung. Die Schwierigkeit der laparoskopischen Chirurgie liegt im Erlernen der Operationstechnik, da diese durch den indirekten Zugang und die indirekte Sicht deutlich anspruchsvoller ist als konventionelle offene Verfahren.

The Medical Imaging Interaction Toolkit (MITK) is a powerful tool for image processing and application development with a long history [1, 2]. It provides an easy-to-use superbuild for a variety of development scenarios and a host of customization options. This very exibility can be intimidating to new users and even more experienced users might miss out on new developments due to the sheer size of the code base. In this tutorial we will give an overview on MITK, theunderlying concepts and the usage of the MITK Workbench. Afterwards we will show the attendees how to start their own software development with MITK and what pitfalls to look out for. Due to time constraints this will be in the form of a development demonstration, instead of a hands-on session. This will be followed by an introduction into more advanced usage scenarios, intended to provide a customized Workbench for special work ows or to conform to branding requirements. We will close with newly finished and ongoing developments and the new opportunities which arise based on them.

Cardiac resynchronization therapy is a treatment option for patients suffering from symptomatic heart failure. The problem with this treatment option is, that 30% to 40% of the patients do not respond. One reason might be the inappropriate placement of the left ventricular lead via the coronary sinus. Therefore, endocardial pacing systems have been developed. Nonetheless, the implantation of these devices requires in addition to the knowledge of the anatomy and scar of the left ventricle (LV), the information of the papillary muscles. As pacing in a papillary muscles may lead to severe problems. To overcome this issue, a fully automatic papillary muscle segmentation in 3D LGE-MRI is presented. First, the left ventricle is initialized using a registration basedapproach, afterwards the short axis view of the LV is estimated. In the next step, the blood pool is segmented. Finally, the papillary muscles are extracted using a threshold based approach. The proposed method was evaluated on six 3D LGE-MRI data sets and were compared to gold standard annotations from clinical experts. This comparison resulted in a Dice coefficient of 0.72.

Mit der vermehrten Verbreitung von Verfahren aus dem Bereich des Maschinellen Lernens in der medizinische Bildverarbeitung wird eine große Menge von akkurat annotierten medizinischen Bilddaten benötigt. Die begrenzten Ressourcen von medizinischen Experten entwickeln sich dabei zum Flaschenhals für das gesamte Forschungsgebiet. Eine neuartige Methode zum Annotieren von Daten im großen Stil, die bereits Einzug in die medizinische Bildverarbeitung erhalten hat, istdas sogenannte Crowdsourcing, welches auf dem Outsourcen kognitiver Aufgaben an anonyme Internetbenutzer basiert. Eine große Herausforderung in diesem Zusammenhang ist die hohe Varianz der Annotationsqualität, die in der Regel durch redundante Aufgabenverteilung gelöst wird. In diesem Beitrag stellen wir einen neuartigen Ansatz zur Bewertung der Annotationsqualität auf Basis von Clickstreams vor. Inspiriert von Verfahren zur Analyse von Benutzerverhalten in sozialen Netzwerken [1] und biometrischen Benutzerauthentifizierung [2] konvertieren wir die Benutzereingaben in einen Vektor von Interaktionsmerkmalen und trainieren einen Regressor, der die Annotationsqualität, repräsentiert durch den DICE Koeffizienten, schätzt. Mehrere Annotationen können so konfidenzgewichtet zu einer finalen Objektsegmentierung fusioniert werden. Unter Verwendung von 20.000 Crowdsegmentierungen auf öffentlich verfügbaren Datensätzen zeigen wir (1) dass unser Verfahren mit weniger als der Hälfte der Kosten dieselbe Qualität im Vergleich zum verbreiteten Majority Voting erreicht und (2) hervorragend auf neue Domänen generalisiert, die nicht in den Trainingsbildern enthalten sind. Durch die Kostenersparnis ist das Anwendungspotential für dieneue Methode hoch.

A temporally consistent segmentation of cardiac structures in spatio-temporal cine MRI sequences is a prerequisite for in-depth analyses of the heart dynamics in clinical practice. Despite its great importance, automated cardiac segmentation is still an open problem, especially for spatio-temporal data due to challenging imaging characteristics, large anatomical heart variability, and diversity of cardiac dynamics. To cope with these challenges, an approach for model-based 4D segmentation of the left and right ventricle in clinical cine MRI sequences is presented in this paper. Central to our approach is a 4D statisticalshape model that accounts for both inter- and intra-patient variability. It is fitted to the spatio-temporal image sequence by applying a computationallyefficient MRF-based discrete optimization approach that uses BRIEF descriptors for image matching. The approach is evaluated on 15 cardiac cine MRI sequences of children and adults with different heart abnormalities. The segmentation results are compared with another effective 4D segmentation technique indicating similar segmentation accuracy but improved coherence and runtime performances.

Verfahren zur automatischen Multi-Organ-Segmentierung in medizinischen Bildvolumina beruhen häufig auf annotierten Daten eines Patientenkollektivs (Atlas) und deren Anpassung z.B. durch zeitintensive nichtlineare Registrierung [1]. Bei der MICCAI 2016 Konferenz [2] stellten wir ein registrierungsfreies Framework für die übertragung von Vorwissen in Form von segmentierten Trainingsdaten auf ungesehene Patienten mit Hilfe eines neuen starken Klassifizierer vor: dieVantage Point Forests (VPF). Ähnlich zu Random Decision Forests (RDF) werden schnelle Berechnungszeiten von wenigen Sekunden erreicht.

We propose an image-based method to estimate the needle axis parameters in 3D ultrasound data although the needle is shown as flat, broad shape due to ultrasound elevation beam-width artifacts in ultrasound volumes. For this, ultrasound volumes are correlated with small 3D sticks of various azimuth and polar angles. The maximal correlation magnitude assigns the orientation of one of the sticks to each voxel position. Voxels of similar angles are clustered and modeled as lines by l₂-norm minimization. The method was applied to 44 3D ultrasound volumes showing a 16G needle which was pushed into pork liver. The estimated location and orientation of the needle axes coincide well with expert defined ”reference” needles, as median (97%-quantile) estimation errors for orientation ≤ 1.24◦ (≤ 5.16◦) and distance between needle tip and axis ≤ 0.53mm (≤ 1.16mm) indicate.

We propose a semi-automatic approach for aorta centerline extraction in contrast-enhanced MRI, making aorta length analysis feasible on large scale. Starting from user-specified start and end regions, we extract the aorta path in between the regions automatically. The extraction is formulated as an optimization problem, seeking for the path that most likely runs central to the aorta. To this end, we exploit that the aorta distinguishes from the surrounding by strong image gradients that point inwards to the aorta’s center due to contrast-enhanced imaging. We also include additional means of manual guidance to resolveerroneous cases. Experiments on data of 19 subjects yielded results that are close to the inter-reader variability. The average distance to the ground truth was 1.89 ± 1.54 mm, while aorta lengths deviated by only 0.66 ± 0.49 %.

In this work, a new method for noise reduction in low dose DSA imaging is presented. The algorithm extends an existing approach using the low rank nature of DSA image series to enable considerable reduction of radiation dose while maintaining low image noise level and preserving spatial resolution and temporal dynamics of the DSA series. The algorithm is based on the singular value decomposition (SVD) using a pixel adaptive approach for the noise reduction. For validation of the method an in vivo animal study is examined

Camera-Based Photoplethysmography is a measuring technique that permits the remote assessment of vital signs by using cameras. The face is the preferred area of measurement (region of interest: ROI) that has to be selected automatically for convenient application. Most works use common face detection algorithm for this purpose. However, these approaches often fail if the face is partly occluded or distorted. In this work, we propose an automatic method for ROI detection and tracking that does not rely on facial features. First, a Bayesian skin classifier was applied. Second, the detected areas were refined and tracked by level set segmentation. We tested our method on videos of 70 patients. The determined ROIs were used for signal extraction and heart rate (HR)estimation. The results showed that our method can detect and track suitable skin regions. We achieved a median HR detection rate of 80% which was only 6% lower than when applying manually defined ROIs.

Im Bereich der biomedizinischen Bildanalyse werden vermehrt öffentliche Wettbewerbe (Challenges) durchgeführt, die den Vergleich von Methoden unter denselben Bedingungen ermöglichen. Ergebnisse aus solchen Challenges gewinnen zur Bewertung von Forschungsresultaten – z.B. im Reviewprozess von Publikationen– immer mehr an Bedeutung. Demgegenüber steht eine mangelnde Qualit ätskontrolle im Challengedesign. Dieser Beitrag beruht auf der Hypothese, dass eine unzureichende Qualitätskontrolle zu einer geringen Aussagekraft der Challengeergebnisse führen kann. Basierend auf dem Validierungsprotokoll von Jannin et al. [1] wurden sämtliche biomedizinischen Challenges des Kollektivs Grand Challenges in Biomedical Image Analysis“ [2] bis zum Jahr 2016 erfasst und systematisch analysiert. Wir präsentieren die Analyseergebnisse hinsichtlich der Vollständigkeit des Reportings und des Einflusses verschiedener Entscheidungen im Challengedesign auf das finale Ranking der Teilnehmer. Unsere Analyse demonstriert die Notwendigkeit einer Qualitätskontrolle, welche dazu beitragen sollte, dass Rankings nachvollziehbar sowie reproduzierbar sind und die Aussagef ähigkeit erhöht wird.

This paper proposes a novel method to deal with overexposure for C-arm CT reconstruction. The proposed method is based on recent progress of one bit compressive sensing (1bit-CS), which is to recover sparse signals from sign measurements. Overexposure could be regarded as a kind of sign information, thus the application of 1bit-CS to overexposure correction in CT reconstruction is expected. This method is evaluated on a phantom and its promising performance implies potential application on clinical data.

C-arm cone-beam CT systems have an increasing popularity in the clinical environment due to their highly flexible scan trajectories.Recent work used these systems to acquire images of the knee joint under weight-bearing conditions. During the scan, the patient is in a standing or in a squatting position and is likely to show involuntary motion, which corrupts image reconstruction. The state-of-the-art fully automatic motion compensation relies on fiducial markers for motion estimation. Due to the not reproducible horizontal trajectory, the system has to be calibrated with a calibration phantom before or after each scan. In this work we present a method to incorporate a self-calibration into the existing motion compensation framework without the need of prior geometric calibration. Quantitative and qualitative evaluations on a numerical phantom as well as clinical data, show superior results compared to the current state-of-the-art method. Moreover, the clinical workflow is improved, as a dedicated system calibration for weight-bearing acquisitions is no longer required.

Robust registration of medical images with missing correspondences caused by pathological structures or anatomical variations is still a challenging problem. This paper presents a robust method for groupwise affine registration based on the RASL algorithm [1] that formulates the registration problem as a sparse and low-rank decomposition. We adapt the RASL algorithm for the alignment of 3D image data and introduce a stochastic optimization scheme to enable the computational tractability. Further, a normalization scheme generates more plausible and unique transformations. In our experiments, the algorithm has been applied to various medical images, and proves its suitability for medical image registration. Especially, the approach shows advantages in the presence of pathologies and outperforms iterative groupwise registration based on ITK. The stochastic optimization scheme generates a significant acceleration allowing for a groupwise affine registration of ten 3D CT images in ∼ 5 minutes on CPU without elaborate optimization.

Within a manifold framework, the interpolation of tomographic image time series is investigated. To this end, the metamorphosis model of a manifold of images is taken into account. Based on a variational time discretization, discrete geodesic paths in this space of images are computed. The space discretization is based on finite elements spanned by tensor product cubic B-splines. An efficient implementation is obtained by utilizing graphics hardware and a proper combination of GPU and CPU computation. First results for time series of optical coherence tomography images of a macular degeneration demonstrate the applicability of this geometric concept.

Layered motion estimation (LME) in X-ray fluoroscopy is a challenging, ill-posed and non-convex problem due to transparency effects and the way the image is defined. Minimizing an energy formulation of layered motion estimation is computationally expensive. For clinical usability of this approach, we propose to use primal-dual optimization parallelized using a graphical processing unit (GPU) to reduce the overall run-time of this algorithm. Experimentally this method is able to substantially reduce target registration error by 70% on manually annotated landmarks on five distinct image sequences compared to the static baseline, similar to prior work on this domain. However, the overall runtime of our method on a conventional GPU is less than 3.3 seconds compared to several minutes for the state of the art. Considering typical framerates of X-ray fluoroscopy devices, this runtime makes the application of layered motion estimation feasible for many clinical workflows.

We propose an automatic approach for early detection of adenocarcinoma in the esophagus. High-definition endoscopic images (50 cancer, 50 Barrett) are partitioned into a dataset containing approximately equal amounts of patches showing cancerous and non-cancerous regions. A deep convolutional neural network is adapted to the data using a transfer learning approach. The final classification of an image is determined by at least one patch, for which the probability being a cancer patch exceeds a given threshold. The model was evaluated with leave one patient out cross-validation. With sensitivity and specificity of 0.94 and 0.88, respectively, our findings improve recently published results on the same image data base considerably. Furthermore, the visualization of the class probabilities of each individual patch indicates, that our approach might be extensible to the segmentation domain.

Glioblastoma segmentation is an important challenge in medical image processing. State of the art methods make use of convolutional neural networks, but generally employ only few layers and small receptive fields, which limits the amount and quality of contextual information available for segmentation. In this publication we use the well known UNet architecture to alleviate these shortcomings. We furthermore show that a sophisticated training scheme that uses dynamic sampling of training data, data augmentation and a class sensitive loss allows training such a complex architecture on relatively few data. A qualitative comparison with the state of the art shows favorable performance of our approach.

Limited-angle computed tomography suffers from missing data in the projection domain, which results in intensity inhomogeneities and streaking artifacts in the image domain. We address both challenges by a two-step deep learning architecture: First, we learn compensation weights that account for the missing data in the projection domain and correct for intensity changes. Second, we formulate an image restoration problem as a variational network to eliminate coherent streaking artifacts. We perform our experiments on realistic data and we achieve superior results for destreaking compared to state-of-the-art non-linear filtering methods in literature. We show that our approach eliminates the need for manual tuning and enables joint optimization of both correction schemes.

Mobile, fast virus detection and classification is of increasing importance in times of epidemic diseases being spread by global traveling and transport. A possible solution is the PAMONO sensor, an optical biological sensor that is able to detect (nanometer-sized) viruses and virus-like particles, utilizing surface plasmon resonance. Captured sensor data is given as image sequences, which can be analyzed by methods from the field of image processing, which is the focus this work. We classify single particles based on their size, using state of the art machine learning techniques, namely convolutional neural networks. This classification allows the measurement of individual particle sizes and the compilation of particle size distributions for a given suspension, which contributes to the goal of classifying different virus types. The classification procedure and estimation of distributions is evaluated using real PAMONO sensor image sequences and particles that simulate viruses. The results show that informative features of the SPR signals can be automatically learned, extracted and used for classification, successfully.

This paper discusses fast pose verification for radiation therapy on a new high-speed radiation therapy device. The PHASER system follows the idea of 4th generation CT imaging and allows fast 360◦ treatment using a steerable electron beam. Doing so, dose delivery is possible in few seconds. A major problem, however, is fast verification of the patient pose during treatment. In this paper, we suggest to use a projection-based approach that can be evaluated quickly and allows an accuracy below 1mm as shown by our simulation study based on planning data from six 4D CT data sets.

Rotational coronary angiography allows for volumetric imaging but requires cardiac and respiratory motion management to achieve meaningful reconstructions. Novel respiratory motion compensation algorithms based on data consistency conditions are applied directly in projection domain and, therefore, overcome the need for uncompensated reconstructions. Earlier, we combined single-frame background subtraction and epipolar consistency conditions to compensate for respiratory motion. In this paper, we show that background subtraction also enables motion estimation via optimization of novel Fourier consistency conditions. The proposed method is evaluated in a numerical phantom study. Compared to the uncompensated case, we found a reduction in residual root-mean-square error of 89% when Fourier consistency conditions were used. The results are promising and encourage experiments on clinical data.

In this paper, we study periodic object motion in computed tomography. Specifically, we investigate the phenomenon that motion may–in a sense–be preserved even in a standard analytical reconstruction that assumes a static object. In fact, these preserved motion patterns reappear in a forward-projection of the allegedly static reconstruction. In numerical simulations abstracting the cardiac anatomy, we show that not only the type of motion, but also the sharpness of the boundary of the moving object affects how much of the motion is preserved.

A new protocol for testing osteosynthesis material postoperatively combining semi-automated segmentation and 3D analysis of surface meshes is proposed. By various steps of transformation and measuring, objective data can be collected. In this study the specifications of a locking plate used for mediocarpal arthrodesis of the wrist were examined. The results show, that union of the lunate, triquetrum, hamate and capitate was achieved and that the plate is comparable to coexisting arthrodesis systems. Additionally, it was shown, that the complications detected correlate to the clinical outcome. In synopsis, this protocol is considered beneficial and should be taken into account in further studies.

Hippocampal segmentation accuracy of out-of-the-box software tools (FreeSurfer, AHEAD, BrainParser) is analysed wrt. potential variability in populations with different pathologies. Findings confirm variabilities wrt. different pathologies but also human rater ground truth and single pathologies exhibit significant variability as well.

In diesem Beitrag wird eine neue Methode zur computerunterst ützten Behandlungsplanung von knöchernen Gesichtsschädelbr üchen unter der Verwendung von Miniplatten vorgestellt. Diese Art von Implantaten wird verwendet, um Knochenbrüche im Gesicht zu behandeln. Nach dem derzeitigen Stand der Technik verwendete Methoden wie die Plattenadaption an stereolithischen Modellen oder auf Basis einer computerunterstützten Planung weisen allerdings eine geringere Flexibilität, Mehrkosten oder hygienische Risiken auf. Mit der hier vorgestellten Software ist es den Chirurgen möglich, das Resultat vorab in nur wenigen Minuten an einem computervisualisierten Modell zu planen und anschließend als STL-Datenformat zu exportieren, um es so in der zukunftsträchtigen 3D-Drucktechnologie verwenden zu können. Dadurch werden Chirurgen in die Lage gesetzt, das generierte Implantat oder eine entsprechende Biegevorlage flexibel für jeden visualisierten Defekt im Behandlungszentrum präzise innerhalb weniger Stunden zu erstellen.

A Prior knowledge of the cochlea’s characteristics helps for selecting the suitable cohlear implants for different patients. Cochlea medical images provide such prior knowledge. Doctors use a manual procedure for image registration and segmentation which is time consuming. The cochlea’s small size and complex structure reveals a big challenge for the automated registration of multi-modal cochlea images.

The development of adenocarcinoma in Barrett’s esophagus is difficult to detect by endoscopic surveillance of patients with signs of dysplasia. Computer assisted diagnosis of endoscopic images (CAD) could therefore be most helpful in the demarcation and classification of neoplastic lesions. In this study we tested the feasibility of a CAD method based on Speeded up Robust Feature Detection (SURF). A given database containing 100 images from 39 patients served as benchmark for feature based classification models. Half of the images had previously been diagnosed by five clinical experts as being ”cancerous”, the other half as ”non-cancerous”. Cancerous image regions had been visibly delineated (masked) by the clinicians. SURF features acquired from full images as well as from masked areas were utilized for the supervised training and testing of an SVM classifier. The predictive accuracy of the developed CAD system is illustrated by sensitivity and specificity values. The results based on full image matching where 0.78 (sensitivity) and 0.82 (specificity) were achieved, while the masked region approach generated results of 0.90 and 0.95, respectively.

Bei der Verbreitung und Übertragung von Infektionen im Krankenhaus sind die Hände ein zentraler Infektionsweg. Die korrekt ausgeführte hygienische Händedesinfektion ist deshalb entscheidend, um nosokomiale Infektionen zu verhindern. Unser Prototyp, aus voran gegangenen Arbeiten, bewertet die Händedesinfektion mit Hilfe von Infrarotthermographie. Ziel ist dem medizinischen Personal im Krankenhausalltag unmittelbare Rückmeldung über die Qualität ihrer Händedesinfektion zu geben. Dazu ist es essentiell die Hände imWärmebild zu detektieren, was bei ähnlicher Umgebungs- und Handoberflächentemperatur nicht möglich ist. Hier stellen wir eine Erweiterung unseres Systems vor, welche die benötigte Handdetektion in einem Farbbild robust ermöglicht und transformieren die Ergebnisse anschließend ins Wärmebild.

Photoakustische Tomografie (PAT) ist eine neue strukturelle und funktionale Bildgebung, welche die Darstellung von optischen Absorbern im Gewebe ermöglicht. Im Gegensatz zu anderen weit verbreiteten Modalitäten ermöglicht PAT die Aufnahme von Bildern in Echtzeit und ohne Strahlungsbelastung. Dabei bietet PAT eine hohe räumliche und zeitliche Auflösung. Anders als etablierte optische Bildgebungsverfahren ist PAT in der Lage, optische Parameter mehrere Zentimeter tief im Gewebe zu messen.

Zur Unterstützung onkologischer Interventionen können durch die Registrierung präoperativer Bildaten zu intraoperativen Cone-Beam-Computertomographieaufnahmen (CBCT) zusätzliche Informationen über die Anatomie und Morphologie des Patienten erhalten werden. In der vorliegenden Arbeit wird eine neuartige Metrik für die gradientenbasierte Bildregistrierung vorgestellt.

This paper discusses a kernel ridge regression (KRR) model for motion estimation in radiotherapy. Using KRR, dense internal motion fields are estimated from high-dimensional surrogates without the need for prior dimensionality reduction. We compare the proposed model to a related approach with dimensionality reduction in the form of principal component analysis and principle component regression. Evaluation was performed in a simulation study based on nine 4D CT patient data sets achieving a mean estimation error of 0.84 ± 0.21mm for our approach.

Manual ischemic stroke lesion segmentation in MR image data is a time-consuming task subject to inter-rater variability. Reliable automated lesion segmentation is of high interest for clinical trials and research in ischemic stroke. However, recent segmentation challenges (e.g. ISLES 2015) illustrate that current state-of-the-art approaches still lack accuracy and ischemic stroke segmentation remains a complicated problem. Within this context, low rank-&-sparse matrix decomposition (also known as robust PCA, RPCA) and RPCA-based non-linear subject-toatlas registration could provide valuable segmentation prior information. The aim of this study is to evaluate the suitability of RPCA and RPCAbased registration for ischemic stroke segmentation in follow-up FLAIR MR data sets. Building on a top-ranked segmentation approach of ISLES 2015, the performance of RPCA sparse component image information as random forest (RF) feature is evaluated. A comprehensive feature-byfeature comparison of the segmentation performance with and without RPCA sparse component information as RF feature illustrate the potential of low rank-&-sparse decomposition to improve stroke segmentation.

In 2014, about 26 million people were suffering from heart failure. Symptomatic heart failure is treated by cardiac resynchronization therapy. However, 30% to 50% do not clinically respond after the implantation of a biventricular pacemaker. To improve the success rate, the quantification of a patient’s scar burden is very important. Lategadolinium- enhanced magnetic resonance imaging is used to visualize regions of scarring in the left ventricle. Scar is very hard to visualize and interpret in 3D. To solve this, an automated scar layer generation method is proposed. The scar is divided into layers and an interactive scrolling is provided. This method allows for precise treatment planning. With the scar layer visualization, eight clinical experts were asked to decide if the scar is epicardial or endocardial. The correct location was identified in 93.75% of the cases using the scar layer visualization.

Digital pathology, driven by the increasing capabilities of modern computers, is an emerging field within medical research and diagnostics. A re-occurring task in pathology is the analysis of immunohistochemical (IHC) stains, i.e. stains in which a specific type of immune cell is highlighted using corresponding antibodies. Automatic quantification of these images is a challenge due to large image sizes of up to 10 gigapixels, but provides a more objective and reproducible evaluation than the exhaustive task of manual analysis. In this context, we compare counting measures against area-based measures in the case of cytoplasmic and membrane-bound IHC stains. Our evaluation indicates a superior performance of the area-based method which reaches a Jaccard index of approximately 80%, while cell nuclei count-based approaches can be severely affected by variance due to masking effects when the cytoplasmic chromogenic staining covers the blue nuclear counterstain

Accurate localization of sets of anatomical landmarks is a challenging task, yet often required in automatic analysis of medical images. Several groups – e.g., Donner et al. – have shown that it is beneficial to incorporate geometrical relations of landmarks into detection procedures for complex anatomical structures. In this paper, we present a two-step approach (compared to three steps as suggested by Donner et al.) combining regression tree ensembles with a Conditional Random Field (CRF), modeling spatial relations. The comparably simple combination achieves a localization rate of 99.6% on a challenging hand radiograph dataset showing high age-related variability, which is slightly superior than state-of-the-art results achieved by Hahmann et al.

2D/3D image fusion is used for a variety of interventional procedures. Overlays of 2D images with perspective-correctly rendered 3D images provide the physicians additional information during the interventions. In this work, a real-time capable 2D/3D registration framework is presented. An adapted parallelization using GPU is investigated for the depth-aware registration algorithm. The GPU hardware architecture is specially taken into account by optimizing memory access patterns and exploiting CUDA-texture memory. The real-time capability is achieved with a median runtime of one 2D/3D registration iteration of 86.1ms with an median accuracy of up to 1.15 mm.

Hemodynamic and mechanical parameters are assumed to play an important role in explaining the initiation, growth and rupture of cerebral aneurysms. Pulsatile deformation of the vascular system due to the changes in pressure during the cardiac cycle are of high interest. Typical spatial and temporal resolution of the image data causes the quantification of geometric deformation to be challenging. In addition, flow velocity changes and the inflow of contrast agents cause vessel intensity variations. These variations in intensity can be mistaken as geometric deformations, leading to an overestimation of the true geometric deformation. In this work, a novel flow phantom is designed to generate ground-truth datasets, which are used to further investigate the relationship between intensity variations and the estimation of geometric deformation. The ground truth image data is used to investigate feasibility and limits of pulsatile deformation estimation using non-linear registration.

For prospective motion correction in ultrahigh-field magnetic resonance imaging, optical tracking is employed by fixating a tailored dental splint on the subject’s upper jaw. However, producing such dental splints is cumbersome and exposes subjects to significant discomfort due to the required dental cast preparation. To catalyze the production of these custom-made splints, we propose a semi-automated workflow.By retrospectively digitizing dental casts of five subjects using a 3D laser scanner, we virtually graft the dental splints and, once visually analyzed and approved, they are fabricated by a 3D printer.This process increases subject comfort and reduces the preparation time.

A-priori knowledge of the number of fibers in a voxel is mandatory and crucial when reconstructing multi-fiber voxels in diffusion MRI. Especially for clinical purposes, this estimation needs to be stable, even when only few gradient directions are acquired. In this work, we propose a novel approach to address this problem based on a deep convolutional neural network (CNN), which is able to identify important gradient directions and can be directly trained on real data. To obtain a ground truth using real data, 100 uncorrelated Human Connectome Project datasets are utilized, with a state-of-the-art framework used for generating a relative ground truth. It is shown that this CNN approach outperforms other state-of-the-art machine learning approaches.

Recent C-arm CT systems allow for the examination of a patient’s knees under weight-bearing conditions. The standing patient tends to show involuntary motion, which introduces motion artifacts in the reconstruction. The state-of-the-art motion correction approach uses fiducial markers placed on the patients’ skin to estimate rigid leg motion. Marker placement is tedious, time consuming and associated with patient discomfort. Further, motion on the skin surface does not reflect the internal bone motion. We propose a purely projection based motion estimation method using consistency conditions of X-ray projections. The epipolar consistency between all pairs of projections is optimized over various motion parameters. We validate our approach by simulating motion from a tracking system in forward projections of clinical data. We visually and numerically assess reconstruction image quality and show an improvement in Structural Similarity from 0.912 for the uncorrected case to 0.943 using the proposed method with a 3D translational motion model. Initial experiments showed promising results encouraging further investigation of practical applicability.

Statistical shape, appearance, and motion models are widely used as priors in medical image analysis to, for example, constrain image segmentation [1] and motion estimation results [2]. These models try to learn a compact parameterization of the space of plausible object instances from a population of observed samples using low-rank matrix approximation methods (SVD or PCA). The quality of these models heavily depends on the quantity and quality of the training population. As it is usually quite challenging to collect large and representative training populations, models used in practice often suffer from a limited expressiveness

Bei circa 11% aller Frakturen des oberen Sprunggelenks (OSG) treten akute Verletzungen der Syndesmose auf, die aufgrund ihrer Instabilität einen operativen Eingriff erfordern [1]. Dabei kann die Stabilisierung mittels Stellschraube zu einer Fehlstellung der Fibula führen, welche ohne Korrektur mit einer Verschlechterung der Lebensqualität des Patienten einhergehen kann. Der Einsatz mobiler 3D C-Bogen ermöglicht eine räumliche Interpretation der Anatomie bei der Verifikation des Repositionsergebnisses. Gleichzeitig stellt die variable Ausrichtung des Scanners zum Patienten eine große Herausforderung beim Vergleich mit anderen Datensätzen dar. Bei der Beurteilung von OSG Frakturen mit Beteiligung der Syndesmose kann ein Vergleich mit der gesunden Gegenseite sinnvoll sein. Da ein weiterer Scan jedoch zusätzliche Strahlenbelastung sowie eine Erhöhung der Operationsdauer bedeutet, sollen stattdessen Einzelprojektionen der gesunden Gegenseite analysiert werden. In der vorliegenden Arbeit wird die Ausrichtung des Datensatzes durch die Detektion des tibiotalaren Gelenkspaltes bestimmt [2]. Ein Quadtree-basierter hierarchischer Varianzvergleich identifiziert potentielle Konturpunkte. Aus diesen werden dann mit Hilfe von Hough Transformationen Schaftkonturen und der Gelenkspalt extrahiert. Die Methode wurde auf 13 C-Bogen Datensätzen mit jeweils 100 Einzelprojektionen angewandt. Dazu wurden die anatomischen Sichtebenen von jeweils drei Unfallchirurgen manuell eingestellt, auf die Einzelprojektionen projiziert und mit den berechneten Ebenen verglichen. Die resultierende Korrelation zwischen Winkelabweichung und korrespondierendem Winkel gibt Aufschluss über bevorzugte Aufnahmerichtungen und dient als Basis für weiterführende klinische Experimente

Thoracic and abdominal multi-organ segmentation has been a challenging problem due to the inter-subject variance of human thoraxes and abdomens as well as the complex 3D intra-subject variance among organs. In this paper, we present a preliminary method for automatically segmenting multiple organs using non-enhanced CT data. The method is based on a simple framework using generic tools and requires no organ-specific prior knowledge. Specifically, we constructed a grayscale CT volume along with a probabilistic atlas consisting of six thoracic and abdominal organs: lungs (left and right), liver, kidneys (left and right) and spleen. A non-rigid mapping between the grayscale CT volume and a new test volume provided the deformation information for mapping the probabilistic atlas to the test CT volume. The evaluation with the 20 VISCERAL non-enhanced CT dataset showed that the proposed method yielded an average Dice coefficient of over 95% for the lungs, over 90% for the liver, as well as around 80% and 70% for the spleen and the kidneys respectively

Every surgical intervention results in physical injuries. Therefore, the patient’s consent is required to avoid liability in case of bodily harm. In a lot of countries a stepwise clarification process is common, combining written and verbal clarification, the latter in form of a conversation between patient and surgeon. However, many studies have shown that the quality of patient information is a weak spot in surgical treatment processes. Our approach, exemplary displayed for minimally invasive spine interventions, supports the clarification conversation by displaying intuitive, comprehensible and interactive 2D and 3D visualisations of both, the patient-specific anatomy and pathologies. Furthermore, information about surgical plans in minimally invasive interventions, like radiofrequency or microwave ablation, could be demonstrated by virtually placing applicators within the patient-customized 3D scene. Visualisation and medical application experts evaluated the contribution and usability of this tool.

Bewegungsdaten in der Strahlentherapie werden als zeitabhängige, dreidimensionale, diskretisierte Vektorfelder beschrieben. Deren Visualisierung stellt eine Herausforderung in den Bereichen Speichermanagement und Berechnungszeit dar. Die Darstellung eines typischen 3D-Vektorfelds mithilfe 3D Pfeil-Glyphen in jedem Voxel bei einer typischen Auflösung würde ~ 10GB Speicher und ~ 30 s CPU Berechnungszeit benötigen. Soll ein Vektorfeld nun auch mittels anderer Darstellungsarten wie z.B. Flusstrajektorien darstellt werden um die übersichtlichkeit zu erhöhen, so vervielfacht sich auch die benötigte Berechnungszeit trotz der Reduktion der Glyphenanzahl. Um zukünftig 4D Lungenatmung animiert darzustellen, müssen Methoden entwickelt werden um Informationen aus einem Vektorfeld zu extrahieren ohne die Balance zwischen Ressourcenaufwand und vollständiger Darstellung zu verlieren.

Interactive segmentation promises to combine the speed of automatic approaches with the reliability of manual techniques. Its performance, however, depends largely on live iterative inputs by a human supervisor. For the task of glioblastoma segmentation in MRI data using a Random Forest pixel classifier we quantify the benefit in terms of speed and segmentation quality of user inputs in falsely classified regions as opposed to guided annotations in regions of high classifier uncertainty. The former results in a significantly higher area under the curve of the Dice score over time in all tumor categories. Exponential fits reveal a significantly higher final Dice score for larger tumor regions (gross tumor volume and edema) but not for smaller regions (necrotic core, non-enhancing abnormalities and contrast-enhancing tumor). Time constants of the exponential fits do not differ significantly.

Das Erstellen einer Trainingsbasis für lernbasierte Methoden zur Gewebecharakterisierung ist häufig fehleranfällig und zeitaufwendig. Sparse and Unambigious (SUR)-Annotationen können hier den Aufwand reduzieren, aber die Annotation mehrerer Gewebeklassen sind dennoch aufwendig. Wir stellen einen Ansatz vor, der das Training ermöglicht, wenn lediglich eine Gewebeklasse annotiert wurde. Diese als positive and unlabeled Learning”(PU-Learning) bezeichnete Methode reduziert den manuellen Annotationsaufwand. In unserer Arbeit zeigen wir zudem, dass die erzielten Segmentierungen sich nicht statistisch signifikante vom Stand der Forschung unterscheiden.

The extraction of microstructures (like microcalcifications or masses) from a DBT phantom can be used for image quality assessment. A complete specification includes exact positions and dimensions of the microstructures, which is not always available. We propose a technique to estimate the required ground truth data from a set of multiple acquisitions. A 3D registration algorithm for DBT data is used to identify different breast phantom components and to perfectly align microstructures within the phantom. The registered data, showing variations of the same ground truth structures, is then combined to an estimate of the microstructures. This approach could be shown to improve the registration result itself, and to enable the determination of the actual parameters of the microstructures.

Grading of tissue based on microscopic images is a common and challenging task. We propose a new method for grading of wholeslide histology images of invasive breast carcinoma, which is based on mitotic cell detection. The method combines a threshold-based attention mechanism and a deep neural network for mitotic cell detection and grading. Our mitotic cell detector is learned from scratch using object centroids. We achieved competitive results in the recent MICCAI TUPAC16 challenge.

Active shape and appearance models that are commonly employed for fast, regularized organ segmentation have several limitations. Here, we adapt the explicit shape regression framework popularized for deformable face alignment to the simultaneous segmentation of lungs, heart and clavicles in X-ray scans. ESR uses data-driven feature learning and combines multiple non-linear regressors in a cascaded manner. We performed extensive experiments and devised appropriate feature ranges, a suitable data augmentation scheme and representative shapes for multi-initialization. With these extensions we obtained new stateof- the-art results for X-ray segmentation outperforming all previous approaches applied to the same dataset and approaching human observer variability with sub-second computation times.

State of the art radiotherapy enables for precise shaping of dose distributions to maintain tumor control while sparing organs at risk. Online adaptive radiotherapy improves that by taking into account inter-fractional changes of the anatomy throughout the treatment course. Especially variations in the posture of head and neck cancer patients cause large deformations in the anatomy, potentially leading to considerable deviations from the planned dose distributions. Adapting the treatment plan to compensate for such changes requires an accurate assessment of anatomical deformations. Commonly used models do not distinguish between bony tissue and soft tissue and thus result in unrealistic deformations of the bones. We propose an articulated patient model for the head and neck region, which inherently preserves the rigidity of the bones while allowing for modelling arbitrary postures of the patient. An articulated skeleton model is assembled from the segmentations of the bones in the planning CT. For an exemplary head and neck cancer patient, 40 individual bones were connected by 45 joints. All the joints were modelled as ball and socket joints, granting rotational mobility around the three body axes. Propagation of motion within the skeletal model is achieved by using an inverse kinematics approach [1]. On a consumer PC, different postures are generated interactively at a rate of ~25 per second. To investigate the advantages and limitations, this model was used to sample images of arbitrary postures of the patient. In addition to the skeletal model, a chainmail-based model was coupled to account for soft tissue deformation. Generation of realistic complex postures was achieved by using only up to 8 user-defined supporting vectors. The segmentations of the bones were generated manually. The resulting sampled images suggest that our model is able to accurately assess different inter-fractional postures of the patient. In combination with its time-efficient computation, it can further be used in the image registration context.

Automatic image segmentation of stroke lesions could be of great importance for aiding the treatment decision. Convolutional neural networks obtain high accuracy for this task at the cost of prohibitive computational demand for time-sensitive clinical scenarios. In this work, we study the use of classical fully-connected neural networks (FC-NN) based on hand-crafted features, which achieve much shorter runtimes. We show that recent advances in optimization and regularization of deep learning can be successfully transferred to FC-NNs to improve the training process and achieve comparable accuracy to random decision forests.

In radiotherapy, fast image-based tracking of tumor motion during radiation is conventionally achieved with template matching approaches on x-ray based projection scans with limited soft-tissue contrast. Emerging MR guided radiotherapy (MRgRT) devices allow for ionization-free imaging during treatment with superior soft tissue contrast. Currently, real-time imaging with MR is only possible for single slice acquisition (2D cine MRI). In this type of acquisition, breathing-induced tumor motion may change the appearance of the target in the scanned plane. This out-of-plane target motion may affect the accuracy of tracking algorithms based on template matching. In this work, an advanced 4D multiple template library approach was developed in order to enable 3D target localization on 2D cine MRI. The image data used in this work consists of: i) training data of 84.7 seconds of parallel sequential cine MRI (4.5Hz, 11 slices, 256x256px size, 1.56mm pixel spacing, 4.5mm slice thickness, 35 repetitions) capturing the target throughout its motion and ii) tracking data of 200 single slices of cine MRI simulating the real-time data provided by an MRgRT device. A retrospective self-navigated phase-based 4D sorting algorithm was developed for 4DMRI reconstruction of the training data from which the phase dependent multiple templates were extracted. These templates were incorporated in a real time multiple-template matching based tracking algorithm which was executed on the tracking data. For evaluation purposes, an automatic target contouring algorithm was developed and manually generated reference data was used. The 4D self-navigated reconstruction algorithm was able to generate 4DMRIs (256x256x31px, isotropic voxel size) from the training data for up to 10 phases of the respiratory cycle. The tracking results reproduced quasi-elliptic 3D target trajectories comparable (mean error 0.76 ± 0.76mm and max error 3.8mm in through-plane direction) to the ones captured with a device-side 4DMRI sequence (acquisition time > 450 seconds). In comparison to the reference data, the mean tracking error in coronal orientation was 0.62 ± 0.91mm in superiorinferior and 0.45 ± 0.80mm in left-right direction. Out-of-plane motion effects in cine MRI can lead to reduced precision of dose administration in beam tracking. With our approach, accurate estimation of the 4D trajectory of the target volume is realized and out-of-plane motion can be assessed.

Registration of high-resolution 3D polarized light images (PLI) to reference blockface images is necessary for the analysis of brain fiber structures. We present an automatic approach for the registration of high-resolution 3D PLI data of histological sections of a human brain that integrates automatic segmentation, rigid registration, and splinebased elastic registration. We have applied the approach to images of the temporal lobe of the human brain and assessed its accuracy and robustness.

Für die Diagnostik von Schilddrüsen(SD)-Erkrankungen werden üblicherweise 2D-Ultraschall(US)-Scanner verwendet. SD Anomalien erfordern jedoch auch Kenntnisse über räumliche Verteilungen von Gewebestrukturen. Spezielle 3DUS- Scanner sind kompliziert in der Handhabung und teuer in der Anschaffung. Daher wurde ein SD Rekonstruktionsverfahren für 2D-Scandaten entwickelt. Eine SD-US-Untersuchung beginnt mit dem Ansetzen des Schallkopfes im Bereich des Brustbeins mit anschließendem Hinführen zum Zungenbein.

In the adaptive immune system, Calcium (Ca2+) is acting as a fundamental on-switch. Fluorescence microscopy is used to study the underlying mechanisms. However, living cells introduce motion and for the analysis of (sub-)cellular Ca2+ activity a precise motion analysis is necessary. We present an image based workflow to detect and analyze cell motion. We evaluate our approach on Jurkat T-cells using cell motion as observed from actual time series of cell images. Results indicate, that our method is able to detect deformation with an error of 0.2222 ± 0.086μm which is in the range of the image resolution, showing that accurate cell deformation detection is possible and feasible.

A patient model is useful for many clinical applications such as patient positioning, device placement, or dose estimation in case of X-ray imaging. A default or a-priori patient model can be estimated using learning based methods trained over a large database. Different methods can be used to estimate such a default model given a restricted number of the input parameters. We investigated different learning based estimation strategies using patient gender, height, and weight as the input to estimate a default patient surface model. We implemented linear regression, an active shape model, kernel principal component analysis and a deep neural network method. These methods are trained on a database containing about 2000 surface meshes. Using linear regression, we obtained a mean vertex error of 20.8±14.7mm for men and 17.8±11.6mm for women, respectively. While the active shape model and kernel PCA method performed better than linear regression, the results also revealed that the deep neural network outperformed all other methods with a mean vertex error of 15.6±9.5mm for male and 14.5±9.3mm for female models.

Kurzfassung. Eine Strahlentherapie von Tumorpatienten erfolgt in der Regel fraktioniert über mehrereWochen. Die Qualität der Therapie hängt entscheidend davon ab, ob sichergestellt werden kann, dass in jeder Fraktion die Strahlung, wie ursprünglich geplant, appliziert wird. Durch Lagerungsungenauigkeiten des Patienten, Organbewegungen, Deformationen, Gewichtsabnahme etc. kann es zu relevanten Veränderungen kommen. Die Bestrahlung erfolgt dann nicht mehr nach dem ursprünglich optimierten Plan. Adaptive Therapieansätze versuchen die Bewegung der Zielstrukturen abzuschätzen und durch geeignete Maßnahmen zu kompensieren. Diese Arbeit verfolgt das Ziel, Organbewegungen während der Therapie mit Hilfe von Ultraschall (US)-Bildgebung zu detektieren. Um sicherzustellen, dass trotz leichter Bewegungen des Patienten der Schallkopf korrekt positioniert bleibt und den Kontakt zur Patientenoberfläche nicht verliert, wird ein Leichtbauroboter eingesetzt. Der Roboter verfügt über eine Sensorik, um Bewegungen der Patientenoberfläche zu erkennen und durch eine Nachführung des Schallkopfes zu kompensieren. Damit können Atembewegungen berücksichtigt werden und der Roboter kann in Gefahrensituationen ausweichen. Die Roboterapplikation wurde in das MITK eingebettet. Der Schallkopf kann während der Therapie über einen Joystick neu positioniert werden. Erste Tests an Probanden zeigen, dass die Ultraschallbilder zur Lagekontrolle eingesetzt werden können.

Kurzfassung. Räumlich-zeitliche 4D-Cine-MRT-Bilddaten werden in der klinischen Praxis zur Untersuchung der Herzbewegung eingesetzt. Um eine automatisierte Verarbeitung dieser Daten durch Segmentierungsoder Registrierungsverfahren zu gewährleisten, ist als erster Schritt üblicherweise die initiale Bestimmung von Position und Orientierung des Herzens notwendig. Hierfür wurden bisher sowohl einfache grauwertbasierte Verfahren als auch lernbasierte Verfahren vorgeschlagen. Da bisher Vergleiche zwischen Verfahren aus diesen beiden Kategorien fehlen, erfolgt in diesem Beitrag ein quantitativer Vergleich zwischen einem klassisches Verfahren basierend auf der Untersuchung von zeitlichen Grauwertvarianzen und einer lernbasierten Hough Forest-Methode zur Detektion von multiplen Landmarken. Die Ergebnisse unserer Evaluation anhand von 10 4D-Cine-MRT-Bilddaten zeigen bezüglich der Initialisierungsgenauigkeit keine signifikanten Unterschiede zwischen beiden Verfahren.

Drosophila melanogaster, commonly known as fruit fly, is an important genetic model organism for fundamental neuroscience research. The central nervous system of the larva consists of about 12,000 neurons. For efficient information retrieval and automated data mining, each patterns needs to be mapped onto a common reference space. Accurate image registration plays a key role here. We investigate standard and multi-channel image registration approaches. Our evaluation shows that multi-channel registration is beneficial in cases where the staining quality is partially compromised and can be complemented by the information contained in an additional reference channel.

Image and data processing plays an increasingly important role in medical research. Improvements in acquisition techniques and collaborations across traditionally separate research fields, such as fusing genomics and radiological information in decision making, lead to an ever increasing, in number as well as in size, amount of data. Managing and sharing this amount of data remains an important area of study. One well-known solution to this problem is the extensible neuroimaging archive toolkit (XNAT) [1], which provides a storage and management solution for large amounts of disparate data. However the processing of data remains largely separate and requires a lot of manual interaction.

Atlases of the human body have many applications, including for instance the analysis of information from patient cohorts to evaluate the distribution of tumours and metastases. We present a 3D Slicer module that simplifies the task of generating a multi-modal atlas from anatomical and functional data. It provides for a simpler evaluation of existing image and verbose patient data by integrating a database that is automatically generated from text files and accompanies the visualization of the atlas volume. The computation of the atlas is a two step process. First, anatomical data is pairwise registered to a reference dataset with an affine initialization and a B-Spline based deformable approach. Second, the computed transformations are applied to anatomical as well as the corresponding functional data to generate both atlases. The module is validated with a publicly available soft tissue sarcoma dataset from The Cancer Imaging Archive. We show that functional data in the atlas volume correlates with the findings from the patient database.

Confocal laser endomicroscopy is a novel imaging technique which provides real-time in vivo examination and histological analysis of tissue during an ongoing endoscopy. We present an automatic classification system that is able to differentiate between healthy and cancerous tissue of the vocal cords. Textural as well as CNN features are encoded using Fisher vectors and vector of locally aggregated descriptors while the classification is performed using random forests and support vector machines. Two experiments are investigated following a leave-onesequence- out cross-validation and a fixed training and test set approach. Classification rates reach up to 87.6% and 81.5 %, respectively.

For deformable image registration in the context of medical applications, the choice of the transformation model determines the level of detail with which the human anatomy can be described. Transformation models, which are based on physical descriptions and allow for detailed modeling, like finite element or massspring models, however, require high computation times. Fast models based on e.g. splines or diffusion, on the other hand, rely on interpolation between landmarks in high contrast regions of the images and lack anatomical correctness, especially in soft tissue. Therefore, we have developed a heterogeneous transformation model based on the chainmail approach [1] with the aim of improving the characteristic deformation behavior of soft tissue within short calculation times.

In 3D digital subtraction angiography, the propagation of iodine-based contrast agent in cerebral vessels implies a delayed enhancement of soft tissue, i.e. parenchyma, which causes inconsistencies across the acquired projection images that impair the quality of the reconstructed volumes. In order to cope with this issue, we perform an estimation of contrast-enhanced parenchyma in projection images. The estimation is based on a vessel segmentation and an iterative interpolation of segmented vessel pixels. The estimated parenchyma is subsequently separated from the projection images. Thus, only contrast-enhanced vessels remain and data inconsistencies due to late-enhancing parenchyma will be reduced. The method is applied to two datasets of cerebral vasculatures. The image series are compared prior and post to parenchyma subtraction. Reconstructed volumes show minor noise in background voxels. An average increase of 37% in signal-to-noise ratio is achieved.

Classification of breast tumors via dynamic contrast-enhanced magnetic resonance imaging is an important task for tumor diagnosis. In this paper, we present an approach for automatic tumor segmentation, feature generation and classification. We apply fuzzy c-means on cooccurrence texture features to generate discriminative features for classification. High-frequency information is removed via discrete wavelet transform and computation is simplified via principal component analysis before extraction. We evaluate our approach using different classification algorithms. Our experimental results show the performances of different classifiers with respect to sensitivity and specificity.

Die Laserphotokoagulation ist eine wirksame thermische Therapie für zahlreiche Netzhauterkrankungen. Schonendere Behandlungsmethoden vermeiden die klassischen, starken Verödungen, indem das Gewebe kontrolliert erwärmt wird. Dadurch sind die Behandlungsareale unsichtbar, wodurch während der Behandlung nicht mehr nachvollziehbar ist, an welchen Stellen behandelt wurde. In dieser Arbeit wird ein Verfahren vorgestellt, das diese nicht sichtbaren Laserbehandlungsareale der Retina zur Orientierungshilfe in Echtzeit erkennt und darstellt. Durch die Verwendung eines featurebasierten Registrierungsansatzes werden die Behandlungsareale zu einem Mosaik zusammengefügt, auf das die Laserspots eingezeichnet werden. Dabei wird jede Position über ein Konturextraktionsverfahren ermittelt. Die Genauigkeit und Robustheit unseres Ansatzes konnte in drei Experimenten gezeigt werden.

In [1], we propose a novel method for instrument tracking in Retinal Microsurgery (RM) which is apt to withstand the challenges of in-vivo RM visual sequences. The proposed approach is a robust closed-loop framework to track and localize the instrument parts in real-time, based on a dual-random forest (RF). First, a tracker employs the pixel intensities in a RF to infer the bounding box around the tool tip. In the second step, this region of interest is forwarded to another RF, which predicts the locations of the tool joints based on HOG features.

The state-of-the-art method of wound assessment is manually performed by clinicians. Such procedure has limited reproducibility and accuracy, large time consumption and high costs. Novel technologies such as laser scanning microscopy, multi-photon microscopy, optical coherence tomography and hyperspectral imaging [1], as well as devices relying on the structured light sensors [2, 3] have limitations due to high costs and may lack portability and availability. The high prevalence of chronic wounds, however, requires inexpensive and portable devices for 3D imaging of skin lesions.

Aktuelle VR-Trainingssimulatoren von Punktionen gehen oft von statischen 3D-Patientendaten aus oder verwenden eine unrealistisch- periodische Animation der Atembewegung. Existierende Methoden zur Modellierung der Atembewegung schätzen personalisierte Atemmodelle, die auch in VR-Trainingssimulatoren verwendet werden können. Für jeden neuen Patienten ist jedoch eine stark belastende bzw. teure 4DDatenakquisition als Vorraussetzung der Modellbildung notwendig. Die hier entwickelte Methodik erlaubt, eine plausible übertragung existierender Atembewegungsmodelle eines Referenzpatienten auf einen neuen statischen Patienten. Dieser kosten- und dosissparende Ansatz wird hier als Proof-of-Concept für das VR-Training im Leberbereich atmender virtueller Patienten gezeigt.

Instrument tracking is a key step for various computer-aidedinterventions in retinal microsurgery. One of the bottlenecks of state-of-the-art template based algorithms is the (re-)initialization during the surgery. We propose an algorithm for robustly detecting the bounding box around the tool tip together with a failure detection of the tracking algorithm. Hereby, the user input dependent algorithm is transformed into a completely automatic framework without the need of an assistant. The performance was compared to two state-of-the-art methods.

This paper presents an automatic selection of viewpoints, forming a camera path, to support the exploration of cerebral aneurysms. Aneurysms bear the risk of rupture with fatal consequences for the patient. For the rupture risk evaluation, a combined investigation of morphological and hemodynamic data is necessary. However, the extensive nature of the time-dependent data complicates the analysis. During exploration, domain experts have to manually determine appropriate views, which can be a tedious and time-consuming process. Our method determines optimal viewpoints automatically based on input data such as wall thickness or pressure. The viewpoint selection is modeled as an optimization problem. Our technique is applied to five data sets and we evaluate the results with two domain experts by conducting informal interviews.

Structural magnetic resonance imaging data are frequently analyzed to reveal morphological changes of the human brain in dementia. Most contemporary imaging biomarkers are scalar values, such as the volume of a structure, and may miss the localized morphological variation of early presymptomatic disease progression. Neuroanatomical shape descriptors, however, can represent complex geometric information of individual anatomical regions and may demonstrate increased sensitivity in association studies.

Statistical shape models learned from a population of training shapes are frequently used as a shape prior. A key problem associated with their training is to provide a representative and large training set of (manual) segmentations. Therefore, models often suffer from the high-dimension-low-sample-size (HDLSS) problem, which limits their expressiveness and directly affects their performance.