Bildverarbeitung für die Medizin 2020

For histopathological tumor assessment, the count of mitotic figures per area is an important part of prognostication. Algorithmic approaches - such as for mitotic figure identification - have significantly improved in recent times, potentially allowing for computer-augmented or fully automatic screening systems in the future. This trend is further supported by whole slide scanning microscopes becoming available in many pathology labs and could soon become a standard imaging tool. For an application in broader fields of such algorithms, the availability of mitotic figure data sets of sufficient size for the respective tissue type and species is an important precondition, that is, however, rarely met. While algorithmic performance climbed steadily for e.g. human mammary carcinoma, thanks to several challenges held in the field, for many tumor types, data sets are not available. In this work, we assess domain transfer of mitotic figure recognition using domain adversarial training on four data sets, two from dogs and two from humans. We were able to show that domain adversarial training considerably improves accuracy when applying mitotic _gure classification learned from the canine on the human data sets (up to +12.8% in accuracy) and is thus a helpful method to transfer knowledge from existing data sets to new tissue types and species.

Segmentation of bacteria in live cell microscopy image sequences is a crucial task to gain insights into molecular processes. A main challenge is that some bacteria strongly change their appearance during the life cycle as response to fluctuations in environmental conditions. We present a novel deep learning method with shape-based weighting of the loss function to accurately segment bacteria during different stages of the life cycle. We evaluate the performance of the method for live cell microscopy images of Bacillus subtilis bacteria with strong changes during the life cycle.

In this paper, we address the problem of retrospective color shading correction. An extension of the established gray-level shading correction algorithm based on signal envelope (SE) estimation to color images is developed using principal color components. Compared to the probably most general shading correction algorithm based on entropy minimization, SE estimation does not need any computationally expensive optimization and thus can be implemented more effciently. We tested our new shading correction scheme on artificial as well as real endoscopic images and observed promising results. Additionally, an indepth analysis of the stop criterion used in the SE estimation algorithm is provided leading to the conclusion that a fixed, user-defined threshold is generally not feasible. Thus, we present new ideas how to develop a non-parametric version of the SE estimation algorithm using entropy.

Prostate cancer (PCa) is the dominating malignant tumor for men worldwide and across all ethnic groups. If a carcinoma is being suspected, e.g. due to blood levels, trans-rectal punch biopsies of the prostate will be accomplished, while in case of higher stages of the disease the complete prostate is being surgically removed (radical prostatectomy). In both cases prostate tissue will be prepared into histological sections on glass microscope slides according to certain laboratory protocols, and is finally microscopically inspected by a trained histopathologist. Even though this method is well established, it can lead to various problems because of objectivity deficiencies. In this paper, we present a proof of concept of using Artificial Neural Networks (ANN) for automatically analyzing prostate cancer tissue and rating its malignancy using tissue microarrays (TMAs) of sampled benign and malignant tissue.

Modern, state-of-the-art deep learning approaches yield human like performance in numerous object detection and classification tasks. The foundation for their success is the availability of training datasets of substantially high quantity, which are expensive to create, especially in the field of medical imaging. Crowdsourcing has been applied to create large datasets for a broad range of disciplines. This study aims to explore the challenges and opportunities of crowd-algorithm collaboration for the object detection task of grading cytology whole slide images. We compared the classical crowdsourcing performance of twenty participants with their results from crowd-algorithm collaboration. All participants performed both modes in random order on the same twenty images. Additionally, we introduced artificial systematic flaws into the precomputed annotations to estimate a bias towards accepting precomputed annotations. We gathered 9524 annotations on 800 images from twenty participants organised into four groups in concordance to their level of expertise with cytology. The crowd-algorithm mode improved on average the participants’ classification accuracy by 7%, the mean average precision by 8% and the inter-observer Fleiss’ kappa score by 20%, and reduced the time spent by 31%. However, two thirds of the artificially modified false labels were not recognised as such by the contributors. This study shows that crowd-algorithm collaboration is a promising new approach to generate large datasets when it is ensured that a carefully designed setup eliminates potential biases.

Deep learning based methods have not reached clinically acceptable results for common medical registration tasks that could be adequately solved using conventional methods. The slower progress compared to image segmentation is due to the lower availability of expert correspondences and the very large learnable parameter space for naive deep learning solutions. We strongly believe that it is necessary and beneficial to integrate conventional optimisation strategies as differentiable modules into deep learning based registration.

Retinal vessel segmentation is an essential step for fundus image analysis. With the recent advances of deep learning technologies, many convolutional neural networks have been applied in this field, including the successful U-Net. In this work, we firstly modify the U-Net with functional blocks aiming to pursue higher performance. The absence of the expected performance boost then lead us to dig into the opposite direction of shrinking the U-Net and exploring the extreme conditions such that its segmentation performance is maintained. Experiment series to simplify the network structure, reduce the network size and restrict the training conditions are designed. Results show that for retinal vessel segmentation on DRIVE database, U-Net does not degenerate until surprisingly acute conditions: one level, one filter in convolutional layers, and one training sample. This experimental discovery is both counter-intuitive and worthwhile. Not only are the extremes of the U-Net explored on a well-studied application, but also one intriguing warning is raised for the research methodology which seeks for marginal performance enhancement regardless of the resource cost.

Chronic obstructive pulmonary disease (COPD) is a lung disease that is not fully reversible and one of the leading causes of morbidity and mortality in the world. Early detection and diagnosis of COPD can increase the survival rate and reduce the risk of COPD progression in patients. Currently, the primary examination tool to diagnose COPD is spirometry. However, computed tomography (CT) is used for detecting symptoms and sub-type classification of COPD. Using different imaging modalities is a diffcult and tedious task even for physicians and is subjective to inter-and intra-observer variations. Hence, developing methods that can automatically classify COPD versus healthy patients is of great interest. In this paper, we propose a 3D deep learning approach to classify COPD and emphysema using volume-wise annotations only. We also demonstrate the impact of transfer learning on the classification of emphysema using knowledge transfer from a pre-trained COPD classification model.

Volumenwachstumssegmentierungstechniken sind oftmals mit der Übersegmentierung angrenzender Organe oder Strukturen behaftet. Künstlich eingebrachte Segmentierungsbarrieren als Nebenbedingungen helfen hierbei. Aktuell werden diese Markierungen häufig noch als manuelle Scribbles vom Benutzer i.d.R. mühsam schichtweise erstellt. Hier wird ein neuer vollautomatischer Ansatz zum Finden von virtuellen 3D-Barrieren mit maschinellen Lernmethoden vorgestellt. Die Abstandsfehler zu Referenzbarrieren liegen zwischen 4,9±1,3 und 10,3±3,6mm.

Creating patient-specific simulation models helps to make customised implant or treatment plans. To create such models, exact locations of the Origin and Insertion Points of the Ligaments (OIPL) are needed. Locating these OIPL is usually done manually and it is a time-consuming procedure.

Reliable recognition and microscopic differentiation of malignant and non-malignant leukocytes from peripheral blood smears is a key task of cytological diagnostics in hematology [1]. Having been practised for well over a century, cytomorphological analysis is still today routinely performed by human examiners using optical microscopes, a process that can be tedious, time-consuming, and suffering from considerable intra-and inter-rater variability [2]. Our work aims to provide a more quantitative and robust decision-aid for the differentiation of single blood cells in general and recognition of blast cells characteristic for Acute Myeloid Leukemia (AML) in particular.

Clinical studies have shown that skeletal muscle mass, sarcopenia and muscle atrophy can be used as predictive indicators for morbidity and mortality after various surgical procedures and in different medical treatment methods. At the same time, the major psoas muscle has been has been used as a tool to assess total muscle volume. From the image processing side it has the advantage of being one of the few muscles that are not surrounded by other muscles at all times, thereby allowing simpler segmentation than in other muscles. The muscle is fully visible on abdominal CT scans, which are for example performed in clinical workups before surgery. Therefore, automatic analysis of the psoas major muscle in routine CT scans would aid in the assessment of sarcopenia without the need for additional scans or examinations. To this end, we present a method for fully automated segmentation of the psoas major muscle in abdominal CT scans using a combination of methods for semantic segmentation and shape analysis. Our method outperforms available approaches for this task, additionally we show a good correlation between muscle volume and population parameters in different clinical datasets.

The locus coeruleus (LC) is a small brain structure in the brainstem that may play an important role in the pathogenesis of Alzheimer’s Disease (AD) and Parkinson’s Disease (PD). The majority of studies to date have relied on using manual segmentation methods to segment the LC, which is time consuming and leads to substantial interindividual variability across raters. Automated segmentation approaches might be less error-prone leading to a higher consistency in Magnetic Resonance Imaging (MRI) contrast assessments of the LC across scans and studies. The objective of this study was to investigate whether a convolutional neural network (CNN)-based automated segmentation method allows for reliably delineating the LC in in vivo MR images. The obtained results indicate performance superior to the inter-rater agreement, i.e. approximately 70% Dice similarity coefficient (DSC).

Analyzing knee cartilage thickness and strain under load can help to further the understanding of the effects of diseases like Osteoarthritis. A precise segmentation of the cartilage is a necessary prerequisite for this analysis. This segmentation task has mainly been addressed in Magnetic Resonance Imaging, and was rarely investigated on contrast-enhanced Computed Tomography, where contrast agent visualizes the border between femoral and tibial cartilage. To overcome the main drawback of manual segmentation, namely its high time investment, we propose to use a 3D Convolutional Neural Network for this task. The presented architecture consists of a V-Net with SeLu activation, and a Tversky loss function. Due to the high imbalance between very few cartilage pixels and many background pixels, a high false positive rate is to be expected. To reduce this rate, the two largest segmented point clouds are extracted using a connected component analysis, since they most likely represent the medial and lateral tibial cartilage surfaces. The resulting segmentations are compared to manual segmentations, and achieve on average a recall of 0.69, which confirms the feasibility of this approach.

WeLineation [1] is a web-based platform supporting scientists of various domains to obtain segmentations, which are close to ground truth (GT) references. A set of image data accompanied by a written task instruction can be uploaded, users can be invited or subscribe to join in. After passing a guided tutorial of pre-segmented example images, users can provide segmentations.

The accurate quantification of visceral and subcutaneous adipose tissue (VAT and SAT) has become a mayor interest worldwide, given that these tissue types represent an important risk factor of metabolic disorders. Currently, the gold standard for measuring volumes of VAT and SAT is the manual segmentation of abdominal fat images from 3D Dixon magnetic resonance (MR) scans – a very expensive and time-consuming process. To this end, we recently proposed Fat-SegNet [1] a fully automated pipeline to accurately segment adipose tissue inside a consistent anatomically defined abdominal region.

Chest X-Ray (CXR) images as part of a non-invasive diagnosis method are commonly used in today’s medical workflow. In traditional methods, physicians usually use their experience to interpret CXR images, however, there is a large interobserver variance. Computer vision may be used as a standard for assisted diagnosis. In this study, we applied an encoder-decoder neural network architecture for automatic lung region detection. We compared a three-class approach (left lung, right lung, background) and a two-class approach (lung, background). The differentiation of left and right lungs as direct result of a semantic segmentation on basis of neural nets rather than post-processing a lung-background segmentation is done here for the first time. Our evaluation was done on the NIH Chest X-ray dataset, from which 1736 images were extracted and manually annotated. We achieved 94:9% mIoU and 92% mIoU as segmentation quality measures for the two-class-model and the three-class-model, respectively. This result is very promising for the segmentation of lung regions having the simultaneous classification of left and right lung in mind.

Model fitting is employed in numerous medical imaging applications for quantitative parameter estimation. Prominent examples include pharmacokinetic modelling of dynamic contrast-enhanced (DCE) MRI data and apparent diffusion coefficient calculations. There are many fitting tools available, however most of them are limited to a special purpose and do not allow for own development and extension.

Optical coherence tomography angiography (OCTA) is an increasingly popular modality for imaging of the retinal vasculature. Repeated optical coherence tomography (OCT) scans of the retina allow the computation of motion contrast to display the retinal vasculature. To the best of our knowledge, we present the first application of compressed sensing for the generation of OCTA volumes. Using a probabilistic signal model for the computation of OCTA volumes and a 3D median filter, it is possible to perform compressed sensing reconstruction of OCTA volumes while suppressing noise. The presented approach was tested on a ground truth, averaged from ten individual OCTA volumes. Average reductions of the mean squared error of 9:67% were achieved when comparing reconstructed OCTA images to the stand-alone application of a 3D median filter.

Elektromagnetisches (EM) Tracking ist beeindflusst von Störeinflüssen durch metallische und EM Materialien im Trackingvolumen, wie sie in vielen Anwendungen in der Medizin vorkommen. Eine Kompensation derartiger Störungen ist insbesondere dann möglich, wenn diese von statischer Natur sind, wie beispielsweise im Fall eines kleinen EM Feldgenerators (FG) der fest mit einer mobilen Bildgebung verbunden ist. In dieser Arbeit wurde eine Vorrichtung zur reproduzierbaren Kalibrierung solcher Aufbauten entwickelt und für den Anwendungsfall eines an einer Ultraschallsonde befestigten FGs experimentell validiert. Mit einer interpolationsbasierten Kalibrierungsmethode zeigte sich eine deutliche, reproduzierbare Reduktion der Feldverzerrung sowohl im Bereich unterhalb der Sonde, als auch in den erstmals untersuchten Seitenbereichen. Der mittlere Fehler der untersuchten 5 cm Distanzen konnte von 1,6mm auf 0,5mm reduziert werden.

Mammography is using low-energy X-rays to screen the human breast and is utilized by radiologists to detect breast cancer. Typically radiologists require a mammogram with impeccable image quality for an accurate diagnosis. In this study, we propose a deep learning method based on Convolutional Neural Networks (CNNs) for mammogram denoising to improve the image quality. We first enhance the noise level and employ Anscombe Transformation (AT) to transform Poisson noise to white Gaussian noise. With this data augmentation, a deep residual network is trained to learn the noise map of the noisy images. We show, that the proposed method can remove not only simulated but also real noise. Furthermore, we also compare our results with state-of-the-art denoising methods, such as BM3D and DNCNN. In an early investigation, we achieved qualitatively better mammogram denoising results.

Endoscopic surgery leads to large amounts of recordings that have to either be stored completely or postprocessed to extract relevant frames. These recordings regularly contain long out-of-body scenes. This paper proposes to apply anomaly detection methods to detect these irrelevant scenes. A conditional generative adversarial networks (GAN) architecture is used to predict future video frames and classify these predictions with an anomaly score. To avoid the successful prediction of anomalous frames due to the good generalization capability of convolutional neural networks (CNNs) we enhance the optimization process with a negative training phase. The experimental results demonstrate promising results for out-of-body sequence detection with the proposed approach. The enhanced GAN training framework can improve the results of the prediction framework by a large margin. The negative training phase reduces the number of false negative (FN) predictions and is shown to counteract a common problem in anomaly detection methods based on convolutional neural networks (CNNs). The good performance in standard metrics also shows the suitability for clinical use.

The extended acquisition time of Magnetic Resonance Imaging (MRI) makes it susceptible to image artifacts caused by subject motion. Artifact presence reduces diagnostic confidence and it could also necessitate a re-scan or an additional examination in extreme cases. Automatic artifact detection at the modality could improve the effciency, reliability and reproducibility of image quality verification. It could also prevent patient recall for additional examination due to unsatisfactory image quality. In this study we evaluate a machine learning method for the automatic detection of motion artifacts in order to instantly recognise problematic acquisitions before the patient has left the scanner. The paper proposes the use of local entropy estimation in the feature extraction stage of the chosen Support Vector Machine (SVM) classifier. Availability of sufficiently large training data set is one of the main constraints in training machine learning models. In order to enable training a model that could detect motion artifacts of varying severity, the paper also proposes a framework for generation of synthetic motion artifatcs in head MRI. On a per-slices basis, the implemented SVM classifier achieved an accuracy of 93.5% in the detection of motion artifacts in clinical MR images.

X-ray dose constantly gains interest in the interventional suite. With dose being generally diffcult to monitor reliably, fast computational methods are desirable. A major drawback of the gold standard based on Monte Carlo (MC) methods is its computational complexity. Besides common variance reduction techniques, filter approaches are often applied to achieve conclusive results within a fraction of time. Inspired by these methods, we propose a novel approach. We down-sample the target volume based on the fraction of mass, simulate the imaging situation, and then revert the down-sampling. To this end, the dose is weighted by the mass energy absorption, up-sampled, and distributed using a guided filter. Eventually, the weighting is inverted resulting in accurate high resolution dose distributions. The approach has the potential to considerably speed-up MC simulations since less photons and boundary checks are necessary. First experiments substantiate these assumptions. We achieve a median accuracy of 96.7% to 97.4% of the dose estimation with the proposed method and a down-sampling factor of 8 and 4, respectively. While maintaining a high accuracy, the proposed method provides for a tenfold speed-up. The overall findings suggest the conclusion that the proposed method has the potential to allow for further effciency.

Additive manufacturing of hydrogel-based implants, as e.g for the human skull are becoming more important as they should allow a modelling of the different natural layers of the skullcap, and support the bone healing process. Nevertheless, the quality, structure and consistency of such 3D-printed hydrogel implants are important for the reliable production, quality assurance and further tests of the implant production. One possibility for non-destructive imaging and quantification of such additive manufactured hydrogels is computed tomography combined with quantitative image analysis. Hence, the goal of this work is the quantitative analysis of the hydrogel-air relationship as well as the automated computation of the hydrogel angles between the different hydrogel layers. This is done by application and evaluation of various classical image analysis methods such as thresholdig, morphological operators, region growing and Fourier transformation of the CT-slices. Results show, that the examined quantities (channels in the hydrogel lattice, angles between the layers) are in the expected ranges, and it can be concluded that the additive manufacturing process may yield usable hydrogel meshes for reconstructive medicine.

Recently, deep learning (DL) found its way to interventional X-ray skin dose estimation. While its performance was found to be acceptable, even more accurate results could be achieved if more data sets were available for training. One possibility is to turn to computed tomography (CT) data sets. Typically, computed tomography (CT) scans can be mapped to tissue labels and mass densities to obtain training data. However, care has to be taken to make sure that the different clinical settings are properly accounted for. First, the interventional environment is characterized by wide variety of table setups that are significantly different from the typical patient tables used in conventional CT. This cannot be ignored, since tables play a crucial role in sound skin dose estimation in an interventional setup, e. g., when the X-ray source is directly underneath a patient (posterior-anterior view). Second, due to interpolation errors, most CT scans do not facilitate a clean segmentation of the skin border. As a solution to these problems, we applied connected component labeling (CCL) and Canny edge detection to (a) robustly separate the patient from the table and (b) to identify the outermost skin layer. Our results show that these extensions enable fully-automatic, generalized pre-processing of CT scans for further simulation of both skin dose and corresponding X-ray projections.

Already before systems malfunction one has to know if hardware components will fail in near future in order to counteract in time. Thus, unplanned downtime is ought to be avoided. In medical imaging, maximizing the system’s uptime is crucial for patients’ health and healthcare provider’s daily business. We aim to predict failures of Head/Neck coils used in Magnetic Resonance Imaging (MRI) by training a statistical model on sequential data collected over time. As image features depend on the coil’s condition, their deviations from the normal range already hint to future failure. Thus, we used image features and their variation over time to predict coil damage. After comparison of different time series classification methods we found Long Short Term Memorys (LSTMs) to achieve the highest F-score of 86.43% and to tell with 98.33% accuracy if hardware should be replaced.

In order to ensure trouble-free operation, prediction of hardware failures is essential. This applies especially to medical systems. Our goal is to determine hardware which needs to be exchanged before failing. In this work, we focus on predicting failures of 20-channel Head/Neck coils using image-related measurements. Thus, we aim to solve a classification problem with two classes, normal and broken coil. To solve this problem, we use data of two different levels. One level refers to one-dimensional features per individual coil channel on which we found a fully connected neural network to perform best. The other data level uses matrices which represent the overall coil condition and feeds a different neural network. We stack the predictions of those two networks and train a Random Forest classifier as the ensemble learner. Thus, combining insights of both trained models improves the prediction results and allows us to determine the coil’s condition with an F-score of 94.14% and an accuracy of 99.09%.

The estimates of traditional segmentation CNNs for the prediction of the follow-up tissue outcome in strokes are not yet accurate enough or capable of properly modeling the growth mechanisms of ischaemic stroke [1]. In our previous shape space interpolation approach [2], the prediction of the follow-up lesion shape has been bounded using core and penumbra segmentation estimates as priors. One of the challenges is to define well-suited growth constraints, as the transition from one to another shape may still result in a very unrealistic spatial evolution of the stroke.

Image-based diagnosis of the human eye is crucial for the early detection of several diseases in ophthalmology. In this work, we investigate the possibility to use image retrieval to support the diagnosis of diabetic retinopathy. To this end, we evaluate different feature learning techniques. In particular, we evaluate the performance of cost functions specialized for metric learning, namely, contrastive loss, triplet loss and histogram loss, and compare them with the classification crossentropy loss. Additionally, we train the network on images graded by diabetic retinopathy severity and transfer the knowledge learned, to retrieve images that are graded by diabetic macular edema severity and evaluate our algorithm on three different datasets. For the task of detecting referable/non-referable diabetic retinopathy, we achieve a sensitivity of 0.84 and specificity of 0.88 on the Kaggle dataset using histogram loss. On the Messidor dataset, we achieve a sensitivity and specificity score of 0.79 and 0.84, respectively.

In this paper, we present a 2D deep multitask learning approach for the segmentation of small structures on the example of avascular necrosis of the femoral head (AVNFH) in MRI. It consists of one joint encoder and three separate decoder branches, each assigned to its own objective. We propose using a reconstruction task to initially pre-train the encoder and shift the objective towards a second necrosis segmentation task in a reconstruction-dependent loss adaptation manner. The third branch deals with the rough localization of the topographical neighborhood of possible femoral necrosis areas. Its output is used to emphasize the roughly approximated location of the segmentation branch’s output. The evaluation of the segmentation performance of our architecture on coronal T1-weighted MRI volumes shows promising improvements compared to a standard U-Net implementation.

For a wide range of clinical applications, such as adaptive treatment planning or intraoperative image update, feature-based deformable registration (FDR) approaches are widely employed because of their simplicity and low computational complexity. FDR algorithms estimate a dense displacement field by interpolating a sparse field, which is given by the established correspondence between selected features. In this paper, we consider the deformation field as a Gaussian Process (GP), whereas the selected features are regarded as prior information on the valid deformations. Using GP, we are able to estimate the both dense displacement field and a corresponding uncertainty map at once. Furthermore, we evaluated the performance of different hyperparameter settings for squared exponential kernels with synthetic, phantom and clinical data respectively. The quantitative comparison shows, GP-based interpolation has performance on par with state-of-the-art B-spline interpolation. The greatest clinical benefit of GP-based interpolation is that it gives a reliable estimate of the mathematical uncertainty of the calculated dense displacement map.

2D-3D registration is central to image guided minimal invasive endovascular therapies such as the treatment of aneurysms. We propose a novel intensity-based 2D-3D registration method based on digitally reconstructed radiographs and the so-called Normalized Gradient Fields (NGF) as a distance measure. We evaluate our method on publicly available clinical data and compare it to five other state-of-the-art 2D-3D registration methods. The results show that our method achieves better accuracy with comparable results in terms of the number of successful registrations and robustness.

High quality reconstruction with interventional C-arm conebeam computed tomography (CBCT) requires exact geometry information. If the geometry information is corrupted, e. g., by unexpected patient or system movement, the measured signal is misplaced in the backprojection operation. With prolonged acquisition times of interventional C-arm CBCT the likelihood of rigid patient motion increases. To adapt the backprojection operation accordingly, a motion estimation strategy is necessary. Recently, a novel learning-based approach was proposed, capable of compensating motions within the acquisition plane. We extend this method by a CBCT consistency constraint, which was proven to be effcient for motions perpendicular to the acquisition plane. By the synergistic combination of these two measures, in and out-plane motion is well detectable, achieving an average artifact suppression of 93 %. This outperforms the entropy-based state-of-the-art autofocus measure which achieves on average an artifact suppression of 54%.

While deep learning has become a methodology of choice in many areas, relatively few deep-learning-based image registration algorithms have been proposed. One reason for this is lack of ground-truth and the large variability of plausible deformations that can align corresponding anatomies. Therefore, the problem is much less constrained than for example image classification or segmentation.

This study investigates a novel data augmentation approach for simulating surgical instruments, tools, and implants by image composition with transformed characters, numerals, and abstract symbols from open-source fonts. We analyse its suitability for the common spatial learning tasks of multi-label segmentation and anatomical landmark detection. The proposed technique is evaluated on 38 clinical intraoperative X-ray images with a high occurrence of objects overlaying the target anatomy. We demonstrate increased robustness towards superimposed surgical objects by incorporating our technique and provide an empirical rationale about the neglectable influence of realistic object shape and intensity information.

The treatment of age-related macular degeneration (AMD) requires continuous eye examinations using optical coherence tomography (OCT). The need for treatment is indicated by the presence or change of disease-specific OCT-based biomarkers. Therapeutic response and recurrence patterns of patients, however, vary widely between individuals and represent a major challenge for physicians.

Recently, Magnetic Resonance Fingerprinting (MRF) was proposed as a quantitative imaging technique for the simultaneous acquisition of tissue parameters such as relaxation times T1 and T2. Although the acquisition is highly accelerated, the state-of-the-art reconstruction suffers from long computation times: Template matching methods are used to find the most similar signal to the measured one by comparing it to pre-simulated signals of possible parameter combinations in a discretized dictionary. Deep learning approaches can overcome this limitation, by providing the direct mapping from the measured signal to the underlying parameters by one forward pass through a network.

Metal artifacts in computed tomography (CT) arise from a mismatch between physics of image formation and idealized assumptions during tomographic reconstruction. These artifacts are particularly strong around metal implants, inhibiting widespread adoption of 3D cone-beam CT (CBCT) despite clear opportunity for intra-operative verification of implant positioning, e. g. in spinal fusion surgery. On synthetic and real data, we demonstrate that much of the artifact can be avoided by acquiring better data for reconstruction in a task-aware and patient-specific manner, and describe the first step towards the envisioned task-aware CBCT protocol.

In computed tomography (CT), data truncation is a common problem. Images reconstructed by the standard filtered back-projection algorithm from truncated data suffer from cupping artifacts inside the field-of-view (FOV), while anatomical structures are severely distorted or missing outside the FOV. Deep learning, particularly the U-Net, has been applied to extend the FOV as a post-processing method. Since image-to-image prediction neglects the data fidelity to measured projection data, incorrect structures, even inside the FOV, might be reconstructed by such an approach. Therefore, generating reconstructed images directly from a post-processing neural network is inadequate. In this work, we propose a data consistent reconstruction method, which utilizes deep learning reconstruction as prior for extrapolating truncated projections and a conventional iterative reconstruction to constrain the reconstruction consistent to measured raw data. Its efficacy is demonstrated in our study, achieving small average root-mean-square error of 24HU inside the FOV and a high structure similarity index of 0.993 for the whole body area on a test patient’s CT data.

Deep learning with convolutional networks (DCNN) has established itself as a powerful tool for a variety of medical imaging tasks. However, DCNNs in particular require strong monitoring by expert annotations, which cannot be generated cost-effectively by laymen. In contrast to manual annotations, the mere availability of medical volume data is not a problem.

Analysing coronary artery plaque segments with respect to their functional significance and therefore their influence to patient management in a non-invasive setup is an important subject of current research. In this work we compare and improve three deep learning algorithms for this task: A 3D recurrent convolutional neural network (RCNN), a 2D multi-view ensemble approach based on texture analysis, and a newly proposed 2.5D approach. Current state of the art methods utilising fluid dynamics based fractional flow reserve (FFR) simulation reach an AUC of up to 0.93 for the task of predicting an abnormal invasive FFR value. For the comparable task of predicting revascularisation decision, we are able to improve the performance in terms of AUC of both existing approaches with the proposed modifications, specifically from 0.80 to 0.90 for the 3D-RCNN, and from 0.85 to 0.90 for the multi-view texture-based ensemble. The newly proposed 2.5D approach achieves comparable results with an AUC of 0.90.

An assumption-free automatic check of medical images for potentially overseen anomalies would be a valuable assistance for a radiologist. Deep learning and especially Variational Auto-Encoders (VAEs) have shown great potential in the unsupervised learning of data distributions. In principle, this allows for such a check and even the localization of parts in the image that are most suspicious.

Assessing coronary artery plaque segments in coronary CT angiography scans is an important task to improve patient management and clinical outcomes, as it can help to decide whether invasive investigation and treatment are necessary. In this work, we present three machine learning approaches capable of performing this task. The first approach is based on radiomics, where a plaque segmentation is used to calculate various shape-, intensity- and texture-based features under different image transformations.

Generative shape models play an important role in medical image analysis. Conventional methods like PCA-based statistical shape models (SSMs) and their various extensions have shown great success modeling natural shape variations in medical images, despite their limitations. Corresponding deep learning-based methods like (variational) autoencoders are well known to overcome many of those limitations. In this work, we compare two conventional and two deep learning-based generative shape modeling approaches to shed light on their limitations and advantages. Experiments on a publicly available 2D chest X-ray data set show that the deep learning methods achieve better specificity and generalization abilities for large training set sizes. However, for smaller training sets, the conventional SSMs are more robust and their latent space is more compact and easier to interpret.

Traditional neuroimage analysis pipelines involve computationally intensive, time-consuming optimization steps, and thus, do not scale well to large cohort studies. With FastSurfer [1] we propose a fast deep-learning based alternative for the automated processing of structural human MRI brain scans, including surface reconstruction and cortical parcellation. FastSurfer consists of an advanced deep learning architecture (FastSurferCNN) used to segment a whole brain MRI into 95 classes in under 1 min, and a surface pipeline building upon this high-quality brain segmentation.

For the characterization of intracranial aneurysms, morphological and hemodynamic parameters provide valuable information. To evaluate these quantities, the separation of the aneurysm from its parent vessel is required by defining a neck curve and the corresponding ostium. A fundamental problem of this concept is the missing ground truth. Recent studies report strong variations for this procedure between medical experts yielding increased interobserver variability for subsequent evaluations. To make further steps towards consensus, we present a web application solution, combining a client based on HTML and JavaScript and a server part utilizing PHP and the Matlab Runtime environment. Within this study, participants are requested to identify the neck curve of five virtual aneurysm models. Furthermore, they can manipulate the ostium surface to model the original parent artery. Our application is now available online and easily accessible for medical experts just requiring an internet browser.

Existing approaches to modeling the dynamics of brain tumor growth, specifically glioma, employ biologically inspired models of cell diffusion, using image data to estimate the associated parameters. In this work, we propose to learn growth dynamics directly from annotated MR image data, without specifying an explicit model, leveraging recent developments in deep generative models. We further assume that imaging is ambiguous with respect to the underlying disease, which is reflcted in our approach in that it doesn’t predict a single growth estimate but instead estimates a distribution of plausible changes for a given tumor.

Spherical coordinate systems have become a standard for analyzing human cortical neuroimaging data. Surface-based signals, such as curvature, folding patterns, functional activations, or estimates of myelination define relevant cortical regions. Surface-based deep learning approaches, however, such as spherical CNNs primarily focus on classification and cannot yet achieve satisfactory accuracy in segmentation tasks. To perform surface-based segmentation of the human cortex, we introduce and evaluate a 2D parameter space approach with view aggregation (p3CNN). We evaluate this network with respect to accuracy and show that it outperforms the spherical CNN by a margin, increasing the average Dice similarity score for cortical segmentation to above 0.9.

In many minimally invasive procedures, image guidance using a C-arm system is utilized. To enhance the guidance, information from pre-operative 3-D images can be overlaid on top of the 2-D fluoroscopy and 2-D/3-D image registration techniques are used to ensure an accurate overlay. Despite decades of research, achieving a highly reliable registration remains challenging. In this paper, we propose a learning-based correspondence estimation, which focuses on contour points and can be used in combination with the point-to-plane correspondence model-based registration. When combined with classical correspondence estimation in a refinement step, the method highly increases the robustness, leading to a capture range of 36mm and a success rate of 98.5%, compared to 14mm and 71.9% for the purely classical approach, while maintaining a high accuracy of 0.430.08mm of mean re-projection distance.

Deformable image registration (DIR) is an important tool in radio therapy where it is used in order to align a baseline CT and daily low-dose cone beam CT (CBCT) scans. DIR allows the propagation of irradiation plans, Hounsfield units and contours of anatomical structures, respectively, which enables tracking of applied doses over time and generation of daily synthetic CT images. Furthermore, DIR allows to overcome segmentation of structures in CBCT images at each fraction.

In minimally invasive procedures, a standard routine of observing the operational site is using image guidance. X-ray fluoroscopy using C-arm systems is widely used. In complex cases, overlays of preoperative 3-D images are necessary to show structures that are not visible in the 2-D X-ray images. The alignment quality may degenerate during an intervention, e. g. due to patient motion, and a new registration needs to be performed. However, a decrease in alignment quality is not always obvious, as the clinician often focuses on structures which are not visible in the 2-D image, and only these structures are visualized in the overlay. In this paper, we propose a learning-based method for detecting different degrees of misalignment. The method is based on point-to-plane correspondences and a pre-trained neural network originally used for detecting good correspondences. The network is extended by a classification branch to detect different levels of misalignment. Compared to simply using the normalized gradient correlation similarity measure as a basis for the decision, we show a highly improved performance, e. g. improving the AUC score from 0.918 to 0.993 for detecting misalignment above 5mm of mean re-projection distance.

Groupwise image registration and the estimation of anatomical shape variation play an important role for dealing with the analysis of large medical image datasets. In this work we adapt the concept of deforming autoencoders that decouples shape and appearance in an unsupervised learning setting, following a deformable template paradigm, and apply its capability for groupwise image alignment. We implement and evaluate this model for the application on medical image data and show its suitability for this domain by training it on middle slice MRI brain scans. Anatomical shape and appearance variation can be modeled by means of splitting a low-dimensional latent code into two parts that serve as inputs for separate appearance and shape decoder networks. We demonstrate the potential of deforming autoencoders to learn meaningful appearance and deformation representations of medical image data.

Analysis of animal locomotion is a commonly used method for analyzing rodent behavior in laboratory animal science. In this context, the open field test is one of the main experiments for assessing treatment effects by analyzing changes in exploratory behavior of laboratory mice and rats. While a number of algorithms for automated analysis of open field experiments has been presented, most of these do not utilize deep learning methods. Therefore, we compare the performance of different deep learning approaches to perform animal localization in open field studies. As our key methodological contribution, we present a novel softargmax-based loss function that can be applied to fully convolutional networks such as the U-Net to allow direct landmark regression from fully convolutional architectures.

Eye movements, blinking and other motion during the acquisition of optical coherence tomography (OCT) can lead to artifacts, when processed to OCT angiography (OCTA) images. Affected scans emerge as high intensity (white) or missing (black) regions, resulting in lost information. The aim of this research is to fill these gaps using a deep generative model for OCT to OCTA image translation relying on a single intact OCT scan. Therefore, a U-Net is trained to extract the angiographic information from OCT patches. At inference, a detection algorithm finds outlier OCTA scans based on their surroundings, which are then replaced by the trained network. We show that generative models can augment the missing scans. The augmented volumes could then be used for 3-D segmentation or increase the diagnostic value.

Point-wise scanning modalities like Optical Coherence Tomography (OCT) or Scanning Laser Ophthalmoscopy suffer from distortions due to the perpetual motion of the eye. While various motion correction approaches have been proposed, the absence of ground truth displacements or images entails a lack of accurate and comparable evaluations. The purpose of this paper is to close this gap by initiating an open source framework for the simulation of realistic eye motion and corresponding artificial distortion of scans, thereby for the first time enabling the community to a) create datasets with accessible ground truth and b) compare the correction of identical motion patterns in data acquired with different scanners or scan patterns. This paper extends previous work on simulation of fixational eye drift via a self-avoiding random walk in a potential to a continuous domain in time and space, allowing the derivation of smooth displacement fields. The model is demonstrated by presenting an examplary motion path, whose properties resemble reported properties of recordings in current literature on fixational eye motion. Furthermore, the artificial distortion of scans is demonstrated by showing a correspondingly distorted image of a virtual raster scan modeled according to the properties of an existing OCT scanner. All experiments can be reproduced and adapted to arbitrary scanner- and raster scan pattern-properties in the publicly available framework. Beyond that, the open source code provides a starting point for the community to integrate extensions like saccadic or axial eye motion.

X-ray imaging is frequently used to facilitate planning and operative guidance for surgical interventions. By capturing patient-specific information prior to and during the procedure, such image-based tools benefit a more reliable and minimally invasive workflow at reduced risk for the patient. To this end, typical assessment involves geometric measurements of patient anatomy, verification of correct positioning of surgical tools and implants, as well as navigational guidance with help of anatomical landmarks and bone morphology.

In endovascular aortic repair (EVAR) procedures fluoroscopy and conventional digital subtraction angiography are currently used to guide the medical instruments inside the patient. Drawbacks of these methods are X-ray exposure and the usage of contrast agents. Moreover, the fluoroscopy provides only a 2D view, which makes the guidance more difficult. For this reason, a catheter prototype including an optical fiber for shape sensing and three electromagnetic (EM) sensors, which provide the position and orientation information, was built to enable a 3D catheter guidance.

Für die Entwicklung eines assistiven Navigationssystems für blinde bzw. sehbehinderte Personen wurde in einem Experiment die Einsetzbarkeit eines vibrotaktilen Armbands als Tool für die Übertragung von Umgebungsinformationen untersucht. Das Tool wurde zunächst für die Navigation der Probanden durch einen Parcours eingesetzt; im Anschluss wurden mit einem Fragenkatalog Angaben zu Tragbarkeit und Einsetzbarkeit erfasst. Die Steuerung der Probanden erfolgte manuell über eine Kontrollapplikation auf einem Smartphone; dies soll in einer späteren Arbeit durch ein System mit Umgebungserkennung für assistive Navigation ersetzt werden. Die Bewegungen und Kollisionen mit Hindernissen im Parcours wurden mit vier Raumkameras erfasst; Durchlaufzeiten und Anzahl Kollisionen wurden ausgewertet. Das Armband wurde auch im Vergleich mit Blindenhund und Blindenstock bewertet. Die Ergebnisse zeigen, dass das Armband schon nach drei Durchgängen deutliche Verkürzungen der Durchlaufzeiten ermöglichte – vergleichbar wie beim geübten Einsatz von einem Blindenstock. Die Tragbarkeit des Armbandes wurde mit Mean Opinions Scores von sehbehinderten Personen als „gut“ bis „exzellent“ eingestuft. Das Armband eignet sich daher insgesamt als Tool für die weitere Entwicklung eines assisitiven Navigationssystems.

The human placenta is vital for the intrauterine growth and development of fetus. It serves several vital functions, including the transmission of nutrients and hormones from the maternal to the fetal circulatory system. During pregnancy, partial infarcts, thrombosis or hemorrhage within the placenta may affect or even reduce the functional regions maintaining the exchange of hormones, oxygen and nutrients with the fetus. This poses a risk to fetal development and should be monitored, since, at a certain point, the nutritious support might not be sufficient anymore. To assess the functional placental tissue, diffusion tensor magnetic resonance imaging (DT-MRI) is used to discriminate different levels of the placental functional state. Highly active regions contain the so-called cotyledons, units that support the fetus with nutrients. In case of their failure, the fetus gets deprived of sufficient nutritious support, which potentially leads to placental intrauterine growth restriction (IUGR). The direct measurement of the functional state of the cotyledons could provide meaningful insight into the current placental energy state. In this paper, we propose a workflow for extracting and visualizing the functional state of a single cotyledon and a combined visualization depicting the energy state of the entire placenta. We provide informal feedback from a radiologist with experience in placental functional data along 17 data sets.

Fundus photography and Optical Coherence Tomography Angiography (OCT-A) are two commonly used modalities in ophthalmic imaging. With the development of deep learning algorithms, fundus image processing, especially retinal vessel segmentation, has been extensively studied. Built upon the known operator theory, interpretable deep network pipelines with well-defined modules have been constructed on fundus images. In this work, we firstly train a modularized network pipeline for the task of retinal vessel segmentation on the fundus database DRIVE. The pretrained preprocessing module from the pipeline is then directly transferred onto OCT-A data for image quality enhancement without further fine-tuning. Output images show that the preprocessing net can balance the contrast, suppress noise and thereby produce vessel trees with improved connectivity in both image modalities. The visual impression is confirmed by an observer study with five OCT-A experts. Statistics of the grades by the experts indicate that the transferred module improves both the image quality and the diagnostic quality. Our work provides an example that modules within network pipelines that are built upon the known operator theory facilitate cross-modality reuse without additional training or transfer learning.

In this paper, we investigate the feasibility of using a standard U-Net for Few-Shot segmentation tasks in very constraint settings. We demonstrate on the example of femur segmentation in X-ray images, that a U-Net architecture only needs few samples to generate accurate segmentations, if the images and the structure of interest only show little variance in appearance and perspective. This is often the case in medical imaging. We also present a novel training strategy for the UNet, leveraging U-Net’s Few-Shot capability for inter-patient consistent protocols. We propose repeatedly enforcing Few-Sample-Fitting the network for faster convergence. The results of our experiments indicate that incrementally fitting the network to an increasing sample set can lead to faster network convergence in constraint few-shot settings.

Generative adversarial networks (GANs) have shown impressive results for photo-realistic image synthesis in the last couple of years. They also offer numerous applications in medical image analysis, such as generating images for data augmentation, image reconstruction and image synthesis for domain adaptation. Despite the undeniable success and the large variety of applications, GANs still struggle to generate images of high resolution.

The current accessibility to large medical datasets for training convolutional neural networks is tremendously high. The associated dataset labels are always considered to be the real “ground truth”. However, the labeling procedures often seem to be inaccurate and many wrong labels are integrated. This may have fatal consequences on the performance of both training and evaluation. In this paper, we show the impact of label noise in the training set on a specific medical problem based on chest X-ray images. With a simple one-class problem, the classification of tuberculosis, we measure the performance on a clean evaluation set when training with label-corrupted data. We develop a method to compete with incorrectly labeled data during training by randomly attacking labels on individual epochs. The network tends to be robust when flipping correct labels for a single epoch and initiates a good step to the optimal minimum on the error surface when flpping noisy labels. On a baseline with an AUC (Area under Curve) score of 0.924, the performance drops to 0.809 when 30% of our training data is misclassified. With our approach the baseline performance could almost be maintained, the performance raised to 0.918.

Quantification of mitotic figures (MF) within the tumor areas of highest mitotic density is the most important prognostic parameter for outcome assessment of many tumor types. However, high intra- and inter-rater variability results from diffculties in individual MF identification and region of interest (ROI) selection due to uneven MF distribution. Deep learning-based algorithms for MF detection and ROI selection are very promising methods to overcome these limitations.

Die Erfolge von Faltungsnetzwerken (Convolutional Neural Networks, CNNs) in der Bildverarbeitung haben in den letzten Jahren große Aufmerksamkeit erregt. Die Erforschung von Verfahren zur Klassifikation von Mimik auf Bildern menschlicher Gesichter stellt in der Medizin eine große Chance für Menschen mit körperlicher Behinderung dar. So können beispielsweise einfach Befehle an einen elektronischen Rollstuhl oder ein Computerprogramm übermittelt werden. Diese Arbeit untersucht, ob und wie weit die Verwendung von Zusatzinformation (hier in Form von Segmentierungen von Gesichtspartien) beim Training eines CNN-Klassifikators die Genauigkeit bezüglich der Entscheidung für verschiedene Kiefer- und Lippenstellungen verbessern kann. Unsere Ergebnisse zeigen, dass die Genauigkeit des CNN-Klassifikators mit dem Detailgrad der verwendeten Segmentierungen zunimmt und außerdem bei Zuhilfenahme von Segmentierungen ein deutlich kleinerer Datensatz (60% der ursprünglichen Datenmenge) ausreicht, um ein ähnlich genaues CNN (im Vgl. zu einem ohne Zusantzinformation) zu trainieren.

Comparison of microvascular circulation on fundoscopic images is a non-invasive clinical indication for the diagnosis and monitoring of diseases, such as diabetes and hypertensions. The differences between intra-patient images can be assessed quantitatively by registering serial acquisitions. Due to the variability of the images (i.e. contrast, luminosity) and the anatomical changes of the retina, the registration of fundus images remains a challenging task. Recently, several deep learning approaches have been proposed to register fundus images in an end-to-end fashion, achieving remarkable results. However, the results are diffcult to interpret and analyze. In this work, we propose an imitation learning framework for the registration of 2D color funduscopic images for a wide range of applications such as disease monitoring, image stitching and super-resolution. We follow a divide-and-conquer approach to improve the interpretability of the proposed network, and analyze both the influence of the input image and the hyperparameters on the registration result. The results show that the proposed registration network reduces the initial target registration error up to 95%.

Advances in medical imaging and convolutional neural networks (CNNs) have made it possible to achieve excellent diagnostic accuracy from CNNs comparable to human raters. However, CNNs are still not implemented in medical trials as they appear as a black box system and their inner workings cannot be properly explained. Therefore, it is essential to assess CNN relevance maps, which highlight regions that primarily contribute to the prediction. This study focuses on the comparison of algorithms for generating heatmaps to visually explain the learned patterns of Alzheimer’s disease (AD) classification. T1-weighted volumetric MRI data were entered into a 3D CNN. Heatmaps were then generated for different visualization methods using the iNNvestigate and keras-vis libraries. The model reached an area under the curve of 0.93 and 0.75 for separating AD dementia patients from controls and patients with amnestic mild cognitive impairment from controls, respectively. Visualizations for the methods deep Taylor decomposition and layer-wise relevance propagation (LRP) showed most reasonable results for individual patients matching expected brain regions. Other methods, such as Grad-CAM and guided backpropagation showed more scattered activations or random areas. For clinically research, deep Taylor decomposition and LRP showed most valuable network activation patterns.

Deep neural networks have achieved tremendous success in various fields including medical image segmentation. However, they have long been criticized for being a black-box, in that interpretation, understanding and correcting architectures is diffcult as there is no general theory for deep neural network design. Previously, precision learning was proposed to fuse deep architectures and traditional approaches.

For minimal invasive procedures like endovascular aortic repair procedures the instruments are navigated with 2D fluoroscopy imaging and digital subtraction angiography, which have several disadvantages. Optical fibers with fiber Bragg gratings (FBG), which allow to sense local strain respectively local curvature and bending angles, can be used for the guidance of medical tools to reduce the X-ray exposure and the used contrast agent. However, FBG-based shape sensing of flexible and long instruments is challenging and the computation includes many steps.

Digital whole-slide images (WSI) are scanned representations of histological tissue at microscopic scale, enabling computer aided diagnosis and remote pathology applications. However, the data sizes may hinder a widespread use, as raw files can easily exceed 10-20GB. In this work, we explore the Scalable High Effciency Video Coding (SHVC) as a replacement for the JPEG standard currently found in most vendor formats. Besides a comparison of the compression rates, this work comprises a user-study to estimate SHVC quantization parameters (QP) and JPEG quality level that threshold the just-noticeable distortions (JND) for a compression below the JND.

Applications in digital histopathology often require costly expert labels to train modern machine learning algorithms. We introduce an adaptation of the Image Quilting algorithm for texture synthesis that is utilized to virtually multiply the tissues and labels. Potential applications are augmentation in neural network training and quality control in intra-rater experiments. We evaluate this method in a subjective expert trial and a quantitative augmented learning scenario.

Percutaneous thermal ablation is a minimally-invasive treatment option for renal cancer. To treat larger tumours, multiple overlapping ablations zones are required. Arrangements with a low number of ablation zones but coverage of the whole tumour volume are challenging to find for physicians. In this work, an open-source software tool with a new planning approach based on the automatic selection from a large number of randomized geometrical arrangements is presented. Two uncertainty parameters are introduced to account for tissue shrinking and tolerance of non-ablated tumour volume. For seven clinical renal T1a, T1b and T2a tumours, ablation plans were proposed by the software. All proposals are comparable to manual plans of an experienced physician with regard to the number of required ablation zones.

Osteoarthritis is a joint disease that commonly affects the hands, feet, spine, as well as the large weight-bearing joints, i. e., the hip, and knees. Worldwide, about 3.6% of the population suffer from osteoarthritis of the knee. If the symptoms are too severe to be treated with medication, the solution is often a total replacement. The precise preoperative planning of implants is a crucial task to achieve a good patient outcome. For planning, usually hybrid 2-D/3-D approaches are used. The main drawback of theses hybrid methods is the different patient positions during the acquisition, namely lying for the 3-D scan and standing for the 2-D scan. We proposed a method that allows acquiring both images in standing positions under natural weight-bearing without having to reposition the patient. To show the feasibility, we provide images from an anthropomorphic leg phantom. A preliminary study with medical experts has shown that the results are promising.

Phantoms for surgical training are able to mimic cutting and suturing properties and patient-individual shape of organs, but lack a realistic visual appearance that captures the heterogeneity of surgical scenes. In order to overcome this in endoscopic approaches, hyperrealistic concepts have been proposed to be used in an augmented reality-setting, which are based on deep image-to-image transformation methods. Such concepts are able to generate realistic representations of phantoms learned from real intraoperative endoscopic sequences.

Assistant physicians typically learn surgical techniques by observation and supervised practice on the patient or using biophantoms. Alternatively, surgical simulators with the possibility of new training possibilities can be used. A number of simulators is already commercially available and might in the future become as important for surgical training as flight simulators. Since so far no simulator is concerned with the training of tracheotomies, a soft-tissue model for simulating tracheotomy was developed. This soft-tissue model is integrated into our ENT surgical simulator for tracheotomy. To model the soft-tissues of the neck (skin and fat), a computed tomography (CT) scan was interactively segmented. For the interaction simulation of a scalpel with the soft-tissue, position based dynamics (PBD) was used, originally developed for the gaming industries. Initial results imply that the proposed approach is able to model soft tissues for virtual surgical training.