VCBM 19: Eurographics Workshop on Visual Computing for Biology and Medicine

Permanent URI for this collection

https://diglib.eg.org/handle/10.2312/2632802

Browse

Now showing 1 - 20 of 28

Colonic Content Assessment from MRI Imaging Using a Semi-automatic Approach
(The Eurographics Association, 2019) Ceballos, Victor; Monclús, Eva; Vázquez, Pere-Pau; Bendezú, Álvaro; Mego, Marianela; Merino, Xavier; Azpiroz, Fernando; Navazo, Isabel; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
The analysis of the morphology and content of the gut is necessary in order to achieve a better understanding of its metabolic and functional activity. Magnetic resonance imaging (MRI) has become an important imaging technique since it is able to visualize soft tissues in an undisturbed bowel using no ionizing radiation. In the last few years, MRI of gastrointestinal function has advanced substantially. However, few studies have focused on the colon, because the analysis of colonic content is time consuming and cumbersome. This paper presents a semi-automatic segmentation tool for the quantitative assessment of the unprepared colon from MRI images. The techniques developed here have been crucial for a number of clinical experiments.
Distance Field Visualization and 2D Abstraction of Vessel Tree Structures with on-the-fly Parameterization
(The Eurographics Association, 2019) Lichtenberg, Nils; Krayer, Bastian; Hansen, Christian; Müller, Stefan; Lawonn, Kai; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
In this paper, we make contributions to the visualization of vascular structures. Based on skeletal input data, we provide a combined 2D and implicit 3D visualization of vasculature, that is parameterized on-the-fly for illustrative visualization. We use an efficient algorithm that creates a distance field volume from triangles and extend it to handle skeletal tree data. Spheretracing this volume allows to visualize the vasculature in a flexible way, without the need to recompute the volume. Illustrative techniques, that have been frequently applied to vascular visualizations often require texture coordinates. Therefore, modifying an object-based algorithm, we propose an image-based, hierarchical optimization process that allows to derive periodic texture coordinates in a frame-coherent way and suits the implicit representation of the vascular structures. In addition to the 3D surface visualization, we propose a simple layout algorithm that applies a 2D parameterization to the skeletal tree nodes. This parameterization can be used to color-code the vasculature or to plot a 2D overview-graph, that highlights the branching topology of the skeleton. We transfer measurements, done in 3D space, to the 2D plot in order to avoid visual clutter and self occlusions in the 3D representation. A visual link between the 3D and 2D views is established via color codes and texture patterns. The potential of our pipeline is shown in several prototypical application scenarios.
DockVis: Visual Analysis of Molecular Docking Data
(The Eurographics Association, 2019) Furmanová, Katarína; Kozlíková, Barbora; Vonásek, Vojtěch; Byška, Jan; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
Molecular docking is one of the key mechanisms for predicting possible interactions between ligands and proteins. This highly complex task can be simulated by several software tools, providing the biochemists with possible ligand trajectories, which have to be subsequently explored and evaluated for their biochemical relevance. This paper focuses on aiding this exploration process by introducing DockVis visual analysis tool. DockVis operates primarily with the multivariate output data from one of the latest available tools for molecular docking, CaverDock. CaverDock output consists of several parameters and properties, which have to be subsequently studied and understood. DockVis was designed in tight collaboration with protein engineers using the CaverDock tool. However, we believe that the concept of DockVis can be extended to any other molecular docking tool providing the users with corresponding computation results.
Evolutionary Pathlines for Blood Flow Exploration in Cerebral Aneurysms
(The Eurographics Association, 2019) Behrendt, Benjamin; Engelke, Wito; Berg, Philipp; Beuing, Oliver; Preim, Bernhard; Hotz, Ingrid; Saalfeld, Sylvia; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
Blood flow simulations play an important role for the understanding of vascular diseases, such as aneurysms. However, analysis of the resulting flow patterns, especially comparisons across patient groups, are challenging. Typically, the hemodynamic analysis relies on trial and error inspection of the flow data based on pathline visualizations and surface renderings. Visualizing too many pathlines at once may obstruct interesting features, e.g., embedded vortices, whereas with too little pathlines, particularities such as flow characteristics in aneurysm blebs might be missed. While filtering and clustering techniques support this task, they require the pre-computation of pathlines densely sampled in the space-time domain. Not only does this become prohibitively expensive for large patient groups, but the results often suffer from undersampling artifacts. In this work, we propose the usage of evolutionary algorithms to reduce the overhead of computing pathlines that do not contribute to the analysis, while simultaneously reducing the undersampling artifacts. Integrated in an interactive framework, it efficiently supports the evaluation of hemodynamics for clinical research and treatment planning in case of cerebral aneurysms. The specification of general optimization criteria for entire patient groups allows the blood flow data to be batch-processed. We present clinical cases to demonstrate the benefits of our approach especially in presence of aneurysm blebs. Furthermore, we conducted an evaluation with four expert neuroradiologists. As a result, we report advantages of our method for treatment planning to underpin its clinical potential.
Feasibility Study For Automatic Bird Tracking and Visualization from Time-Dependent Marine Radar Imagery
(The Eurographics Association, 2019) Ganglberger, Florian; Bühler, Katja; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
In recent years, radar technology has increasingly been used for the monitoring of bird migration. Marine radars are often utilized for this purpose because of their wide accessibility, range, and resolution. They allow the tracking of birds even at night-when most bird migration takes place-over extended periods of time. This creates a wealth of radar images, for which manual annotation of bird tracks is not feasible. We propose a tool for automatic bird tracking and visualization from marine radar imagery. For this purpose, we developed a bird tracking algorithm for vertically recorded radar images that is able to extract quantitative parameters including flight direction, height, and duration. The results can be qualitatively verified by a visualization design that enables domain experts the time-dependent visualization of bird tracks. Furthermore, it allows a preprocessing of radar images taken by screen capturing for device independence. Our tool was used in an ornithological monitoring study to analyze over 200.000 vertically recorded radar images taken in multiple observation periods and locations.
HIFUpm: a Visual Environment to Plan and Monitor High Intensity Focused Ultrasound Treatments
(The Eurographics Association, 2019) Modena, Daniela; Bassano, Davide; Elevelt, Aaldert; Baragona, Marco; Hilbers, Peter A. J.; Westenberg, Michel A.; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
High Intensity Focused Ultrasound (HIFU) is a non invasive therapeutic method, which has been a subject of interest for the treatment of various kinds of tumors. Despite the numerous advantages, HIFU techniques do not reach the high delivery precision like other therapies (e.g., radiotherapy). For this reason, a correct therapy planning and monitoring in HIFU treatments remains a challenge. We propose HIFUpm, a visual analytics approach which enables the visualization of the HIFU simulation results, while guiding the user in the evaluation of the procedure. We illustrate the use of HIFUpm for an ablative treatment of an osteoid osteoma. This use case demonstrates that HIFUpm provides a flexible visual environment to plan and monitor HIFU procedures.
Hybrid Visualization of Protein-Lipid and Protein-Protein Interaction
(The Eurographics Association, 2019) Alharbi, Naif; Krone, Michael; Chavent, Matthieu; Laramee, Robert S.; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
In the Molecular Dynamics (MD) visualization literature, different approaches are utilized to study protein-lipid interactions (PLI) and protein-protein interaction (PPI) in decoupled contexts. However, the two types of interaction occur in the same space-time domain. It is beneficial to study the PLI and PPI in a unified context. Nevertheless, the simulation's size, length, and complexity increase the challenge of understanding the dynamic behavior. We propose a novel framework consisting of four linked views, a time-dependent 3D view, a novel hybrid view, a clustering timeline, and a details-on-demand window. We introduce a selection of visual designs to convey the behavior of PLI and PPI through a unified coordinate system. Abstraction is used to present proteins in hybrid 2D space, a projected tiled space is used to present both PLI and PPI at the particle level in a heat-map style visual design while glyphs are used to represent PPI at the molecular level. We couple visually separable visual designs in a unified coordinate space. The result lets the user study both PLI and PPI separately or together in a unified visual analysis framework. We also exemplify its use with case studies focusing on protein clustering and we report domain expert feedback.
InkVis: A High-Particle-Count Approach for Visualization of Phase-Contrast Magnetic Resonance Imaging Data
(The Eurographics Association, 2019) de Hoon, Niels; Lawonn, Kai; Jalba, Andrei; Eisemann, Elmar; Vilanova, Anna; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
Phase-Contrast Magnetic Resonance Imaging (PC-MRI) measures volumetric and time-varying blood flow data, unsurpassed in quality and completeness. Such blood-flow data have been shown to have the potential to improve both diagnosis and risk assessment of cardiovascular diseases (CVDs) uniquely. Typically PC-MRI data is visualized using stream- or pathlines. However, time-varying aspects of the data, e.g., vortex shedding, breakdown, and formation, are not sufficiently captured by these visualization techniques. Experimental flow visualization techniques introduce a visible medium, like smoke or dye, to visualize flow aspects including time-varying aspects. We propose a framework that mimics such experimental techniques by using a high number of particles. The framework offers great flexibility which allows for various visualization approaches. These include common traditional flow visualizations, but also streak visualizations to show the temporal aspects, and uncertainty visualizations. Moreover, these patient-specific measurements suffer from noise artifacts and a coarse resolution, causing uncertainty. Traditional flow visualizations neglect uncertainty and, therefore, may give a false sense of certainty, which can mislead the user yielding incorrect decisions. Previously, the domain experts had no means to visualize the effect of the uncertainty in the data. Our framework has been adopted by domain experts to visualize the vortices present in the sinuses of the aorta root showing the potential of the framework. Furthermore, an evaluation among domain experts indicated that having the option to visualize the uncertainty contributed to their confidence on the analysis.
Interactive CPU-based Ray Tracing of Solvent Excluded Surfaces
(The Eurographics Association, 2019) Rau, Tobias; Zahn, Sebastian; Krone, Michael; Reina, Guido; Ertl, Thomas; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
Depictions of molecular surfaces such as the Solvent Excluded Surface (SES) can provide crucial insight into functional molecular properties, such as the molecule's potential to react. The interactive visualization of single and multiple molecule surfaces is essential for the data analysis by domain experts. Nowadays, the SES can be rendered at high frame rates using shader-based ray casting on the GPU. However, rendering large molecules or larger molecule complexes requires large amounts of memory that has the potential to exceed the memory limitations of current hardware. Here we show that rendering using CPU ray tracing also reaches interactive frame rates without hard limitations to memory. In our results large molecule complexes can be rendered with only the precomputation of each individual SES, and no further involved representation or transformation. Additionally, we provide advanced visualization techniques like ambient occlusion opacity mapping (AOOM) to enhance the comprehensibility of the molecular structure. CPU ray tracing not only provides very high image quality and global illumination, which is beneficial for the perception of spatial structures, it also opens up the possibility to visualize larger data sets and to render on any HPC cluster. Our results demonstrate that simple instancing of geometry keeps the memory consumption for rendering large molecule complexes low, so the examination of much larger data is also possible.
Interactive Exploded Views for Molecular Structures
(The Eurographics Association, 2019) Sbardellati, Maximilian; Miao, Haichao; WU, Hsiang-Yun; Groeller, Eduard; Barisic, Ivan; Viola, Ivan; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
We propose an approach to interactively create exploded views of molecular structures with the goal to help domain experts in their design process and provide them with a meaningful visual representation of component relationships. Exploded views are excellently suited to manage visual occlusion of structure components, which is one of the main challenges when visualizing complex 3D data. In this paper, we discuss four key parameters of an exploded view: explosion distance, direction, order, and the selection of explosion components. We propose two strategies, namely the structure-derived exploded view and the interactive free-form exploded view, for computing these four parameters systematically. The first strategy allows scientists to automatically create exploded views by computing the parameters from the given object structures. The second strategy further supports them to design and customize detailed explosion paths through user interaction. Our approach features the possibility to animate exploded views, to incorporate ease functions into these animations and to display the explosion path of components via arrows. Finally, we demonstrate three use cases with various challenges that we investigated in collaboration with a domain scientist. Our approach, therefore, provides interesting new ways of investigating and presenting the design layout and composition of complex molecular structures.
Interactive Formation of Statistical Hypotheses in Diffusion Tensor Imaging
(The Eurographics Association, 2019) Abbasloo, Amin; Wiens, Vitalis; Schmidt-Wilcke, Tobias; Sundgren, Pia; Klein, Reinhard; Schultz, Thomas; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
When Diffusion Tensor Imaging (DTI) is used in clinical studies, statistical hypothesis testing is the standard approach to establish significant differences between groups, such as patients and healthy controls. However, diffusion tensors contain six degrees of freedom, and the most commonly used univariate tests reduce them to a single scalar, such as Fractional Anisotropy. Multivariate tests that account for the full tensor information have been developed, but have not been widely adopted in practice. Based on analyzing the limitations of existing univariate and multivariate tests, we argue that it is beneficial to use a more flexible, steerable test. Therefore, we introduce a test that can be customized to include any subset of tensor attributes that are relevant to the analysis task at hand. We also present a visual analytics system that supports the exploratory task of customizing it to a specific scenario. Our system closely integrates quantitative analysis with suitable visualizations. It links spatial and abstract views to reveal clusters of strong differences, to relate them to the affected anatomical structures, and to visually compare the results of different tests. A use case is presented in which our system leads to the formation of several new hypotheses about the effects of systemic lupus erythematosus on water diffusion in the brain.
Layer-Aware iOCT Volume Rendering for Retinal Surgery
(The Eurographics Association, 2019) Weiss, Jakob; Eck, Ulrich; Nasseri, Muhamad Ali; Maier, Mathias; Eslami, Abouzar; Navab, Nassir; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
Retinal microsurgery is one of the most challenging types of surgery, yet in practice, intraoperative digital assistance is rare. The introduction of fast, microscope integrated Optical Coherence Tomography (iOCT) has enabled intraoperative imaging of subsurface structures. However, effective intraoperative visualization of this data poses a challenging problem both in terms of performance and engineering as well as for creating easily interpretable visualizations of this data. Most existing research focuses on visualization of diagnostic OCT data where imaging quality is higher and processing times are not an issue. We introduce a perceptually linear color map for a separated encoding of tissue reflectivity and positional information as chrominance and luminance. Based on this color mapping, we propose a Direct Volume Rendering (DVR) method that aids structure perception. To aid subretinal injection tasks, we introduce a novel Layer-Adjusted Maximum Intensity Projection (LA-MIP), correcting for the natural curvature of the retinal tissue. Expert feedback suggests our methods are preferred over baseline methods. Further research is needed to confirm the benefits of our approach in routine clinical applications.
Medical Animations: A Survey and a Research Agenda
(The Eurographics Association, 2019) Preim, Bernhard; Meuschke, Monique; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
Animation is a potentially powerful instrument to convey complex information with movements, smooth transitions between different states that employ the strong human capabilities to perceive and interpret motion. Animation is a natural choice to display time-dependent data where the dynamic nature of the data is mapped to a kind of video (temporal animation). Clipping planes may be smoothly translated and object transparency adapted to control visibility and further support emphasis of spatial relations, e.g. around a tumor. Animation, however, may also be employed for static data, e.g. to move a camera along a predefined path to convey complex anatomical structures. Virtual endoscopy, where the virtual camera is moved inside an air-filled or fluid-filled structure is a prominent example for these non-temporal animations. Animations, however, are complex visualizations that may depict a larger number of changes in a short period of time. Thus, they need to be assessed in their capability to actually convey information. In this paper, we give a survey of temporal and non-temporal animated visualizations focussed on medical applications and discuss the research potential that arises. To be employed more widely, cognitive limitations, e.g. change blindness, need to be considered. The reduction of complexity in temporal animations is an essential topic to enable the detection and interpretation of changes. Emphasis techniques may guide the user's attention and improve the perception of essential features. Finally, interaction beyond the typical video recorder functionality is considered. Although our focus is medicine, the discussion of a research agenda is partially based on cartography, where animation is widely used.
MedUse: A Visual Analysis Tool for Medication Use Data in the ABCD Study
(The Eurographics Association, 2019) Bartsch, Hauke; Garrison, Laura; Bruckner, Stefan; Wang, Ariel; Tapert, Susan F.; Grüner, Renate; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
The RxNorm vocabulary is a yearly-published biomedical resource providing normalized names for medications. It is used to capture medication use in the Adolescent Brain Cognitive Development (ABCD) study, an active and publicly available longitudinal research study following 11,800 children over 10 years. In this work, we present medUse, a visual tool allowing researchers to explore and analyze the relationship of drug category to cognitive or imaging derived measures using ABCD study data. Our tool provides position-based context for tree traversal and selection granularity of both study participants and drug category. Developed as part of the Data Exploration and Analysis Portal (DEAP), medUse is available to more than 600 ABCD researchers world-wide. By integrating medUse into an actively used research product we are able to reach a wide audience and increase the practical relevance of visualization for the biomedical field.
Molecular Sombreros: Abstract Visualization of Binding Sites within Proteins
(The Eurographics Association, 2019) Schatz, Karsten; Krone, Michael; Bauer, Tabea L.; Ferrario, Valerio; Pleiss, Jürgen; Ertl, Thomas; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
We present a novel abstract visualization for the binding sites of proteins. Binding sites play an essential role in enzymatic reactions and are, thus, often investigated in structural biology. They are typically located within cavities. The shape and properties of the cavity influence whether and how easily a substrate can reach the active site where the reaction is triggered. Molecular surface visualizations can help to analyze the accessibility of binding sites, but are typically prone to visual clutter. Our novel abstract visualization shows the cavity containing the binding site as well as the surface region directly surrounding the cavity entrance in a simplified manner. The resulting visualization resembles a hat, where the brim depicts the surrounding surface region and the crown the cavity. Hence, we dubbed our abstraction Molecular Sombrero, using the Spanish term for 'hat'. Our abstraction is less cluttered than traditional molecular surface visualizations. It highlights important parameters, like cavity diameter, by mapping them to the shape of the sombrero. The visual abstraction also facilitates an easy side-by-side comparison of different data sets. We show the applicability of our Molecular Sombreros to different real-world use cases.
Multiparametric Magnetic Resonance Image Synthesis using Generative Adversarial Networks
(The Eurographics Association, 2019) Haarburger, Christoph; Horst, Nicolas; Truhn, Daniel; Broeckmann, Mirjam; Schrading, Simone; Kuhl, Christiane; Merhof, Dorit; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
Generative adversarial networks have been shown to alleviate the problem of limited training data for supervised learning problems in medical image computing. However, most generative models for medical images focus on image-to-image translation rather than de novo image synthesis. In many clinical applications, image acquisition is multiparametric, i.e. includes contrast-enchanced or diffusion-weighted imaging. We present a generative adversarial network that synthesizes a sequence of temporally consistent contrast-enhanced breast MR image patches. Performance is evaluated quantitatively using the Fréchet Inception Distance, achieving a minimum FID of 21.03. Moreover, a qualitative human reader test shows that even a radiologist cannot differentiate between real and fake images easily.
Pelvis Runner: Visualizing Pelvic Organ Variability in a Cohort of Radiotherapy Patients
(The Eurographics Association, 2019) Grossmann, Nicolas; Casares-Magaz, Oscar; Muren, Ludvig Paul; Moiseenko, Vitali; Einck, John P.; Gröller, Eduard; Raidou, Renata Georgia; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
In radiation therapy, anatomical changes in the patient might lead to deviations between the planned and delivered dose- including inadequate tumor coverage, and overradiation of healthy tissues. Exploring and analyzing anatomical changes throughout the entire treatment period can help clinical researchers to design appropriate treatment strategies, while identifying patients that are more prone to radiation-induced toxicity. We present the Pelvis Runner, a novel application for exploring the variability of segmented pelvic organs in multiple patients, across the entire radiation therapy treatment process. Our application addresses (i) the global exploration and analysis of pelvic organ shape variability in an abstracted tabular view and (ii) the local exploration and analysis thereof in anatomical 2D/3D views, where comparative and ensemble visualizations are integrated. The workflow is based on available retrospective cohort data, which incorporate segmentations of the bladder, the prostate, and the rectum through the entire radiation therapy process. The Pelvis Runner is applied to four usage scenarios, which were conducted with two clinical researchers, i.e., medical physicists. Our application provides clinical researchers with promising support in demonstrating the significance of treatment plan adaptation to anatomical changes.
preha: Establishing Precision Rehabilitation with Visual Analytics
(The Eurographics Association, 2019) Bernold, Georg; Matkovic, Kresimir; Gröller, Eduard; Raidou, Renata Georgia; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
This design study paper describes preha, a novel visual analytics application in the field of in-patient rehabilitation. We conducted extensive interviews with the intended users, i.e., engineers and clinical rehabilitation experts, to determine specific requirements of their analytical process.We identified nine tasks, for which suitable solutions have been designed and developed in the flexible environment of kibana. Our application is used to analyze existing rehabilitation data from a large cohort of 46,000 patients, and it is the first integrated solution of its kind. It incorporates functionalities for data preprocessing (profiling, wrangling and cleansing), storage, visualization, and predictive analysis on the basis of retrospective outcomes. A positive feedback from the first evaluation with domain experts indicates the usefulness of the newly proposed approach and represents a solid foundation for the introduction of visual analytics to the rehabilitation domain.
Robustness Evaluation of CFD Simulations to Mesh Deformation
(The Eurographics Association, 2019) Scheid-Rehder, Alexander; Lawonn, Kai; Meuschke, Monique; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
CFD simulations are an increasingly important method for the non-invasive analysis of risk factors for aneurysm rupture. Their robustness, however, has to be examined more thoroughly before clinical use is possible. We present a novel framework that enables robustness evaluation of CFD simulation according to mesh deformation on patient-specific blood vessel geometry. Our tool offers a guided workflow to generate, run, and visualize OpenFOAM simulations, which significantly decreases the usual overhead of CFD simulations with OpenFOAM. Besides, the deformation of the original geometry allows the user to evaluate the robustness of the simulation without the need to repeat expensive operations of the data pre-processing phase. We assessed the robustness of CFD simulations by applying our framework to several aneurysm data sets.
Semantic Segmentation of Brain Tumors in MRI Data Without any Labels
(The Eurographics Association, 2019) Weninger, Leon; Krauhausen, Imke; Merhof, Dorit; Kozlíková, Barbora and Linsen, Lars and Vázquez, Pere-Pau and Lawonn, Kai and Raidou, Renata Georgia
Brain MR images are one of the most important instruments for diagnosing neurological disorders such as tumors, infections or trauma. In particular, grade I-IV brain tumors are a well-studied subject for supervised deep learning approaches. However, for a clinical use of these approaches, a very large annotated database that covers all of the occurring variance is necessary. As MR scanners are not quantitative, it is unclear how good supervised approaches, trained on a specific database, will actually perform on a new set of images that may stem from a yet other scanner. We propose a new method for brain tumor segmentation, that can not only identify abnormal regions, but can also delineate brain tumors into three characteristic radiological areas: The edema, the enhancing core, and the non-enhancing and necrotic tissue. Our concept is based on FLAIR and T1CE MRI sequences, where abnormalities are detected with a variational autoencoder trained on healthy examples. The detected areas are finally postprocessed via Gaussian Mixture Models and finally classified according to the three defined labels. We show results on the BraTS2018 dataset and compare these to previously published unsupervised segmentation results as well as to the results of the BraTS challenge 2018. Our developed unsupervised anomaly detection approach is on par with previously published methods. Meanwhile, the semantic segmentation - a new and unique model - shows encouraging results.

Browse

Browsing VCBM 19: Eurographics Workshop on Visual Computing for Biology and Medicine by Title

Results Per Page

Sort Options