Group-equivariant convolutional neural networks (G-CNN) heavily rely on parameter sharing to increase CNN's data efficiency and performance. However, the parameter-sharing strategy greatly increases the computational burden for each added parameter, which hampers its application to deep neural network models. In this paper, we address these problems by proposing a non-parameter-sharing approach for group equivariant neural networks. The proposed methods adaptively aggregate a diverse range of filters by a weighted sum of stochastically augmented decomposed filters. We give theoretical proof about how the group equivariance can be achieved by our methods. Our method applies to both continuous and discrete groups, where the augmentation is implemented using Monte Carlo sampling and bootstrap resampling, respectively. Our methods also serve as an efficient extension of standard CNN. The experiments show that our method outperforms parameter-sharing group equivariant networks and enhances the performance of standard CNNs in image classification and denoising tasks, by using suitable filter bases to build efficient lightweight networks.

In this work, we report the set-up and results of the Liver Tumor Segmentation Benchmark (LiTS), which was organized in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI) 2017 and the International Conferences on Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2017 and 2018. The image dataset is diverse and contains primary and secondary tumors with varied sizes and appearances with various lesion-to-background levels (hyper-/hypo-dense), created in collaboration with seven hospitals and research institutions. Seventy-five submitted liver and liver tumor segmentation algorithms were trained on a set of 131 computed tomography (CT) volumes and were tested on 70 unseen test images acquired from different patients. We found that not a single algorithm performed best for both liver and liver tumors in the three events. The best liver segmentation algorithm achieved a Dice score of 0.963, whereas, for tumor segmentation, the best algorithms achieved Dices scores of 0.674 (ISBI 2017), 0.702 (MICCAI 2017), and 0.739 (MICCAI 2018). Retrospectively, we performed additional analysis on liver tumor detection and revealed that not all top-performing segmentation algorithms worked well for tumor detection. The best liver tumor detection method achieved a lesion-wise recall of 0.458 (ISBI 2017), 0.515 (MICCAI 2017), and 0.554 (MICCAI 2018), indicating the need for further research. LiTS remains an active benchmark and resource for research, e.g., contributing the liver-related segmentation tasks in http://medicaldecathlon.com/. In addition, both data and online evaluation are accessible via https://competitions.codalab.org/competitions/17094.

In this paper, we present MetaStore, a metadata management framework for scientific data repositories. Scientific experiments are generating a deluge of data and metadata. Metadata is critical for scientific research, as it enables discovering, analysing, reusing, and sharing of scientific data. Moreover, metadata produced by scientific experiments is heterogeneous and subject to frequent changes, demanding a flexible data model. Currently, there does not exist an adaptive and a generic solution that is capable of handling heterogeneous metadata models. To address this challenge, we present MetaStore, an adaptive metadata management framework based on a NoSQL database. To handle heterogeneous metadata models and standards, the MetaStore automatically generates the necessary software code (services) and extends the functionality of the framework. To leverage the functionality of NoSQL databases, the MetaStore framework allows full-text search over metadata through automated creation of indexes. Finally, a dedicated REST service is provided for efficient harvesting (sharing) of metadata using the METS metadata standard over the OAI-PMH protocol.

Localization microscopy (LM) allows to acquire pointillistic superresolution images of biological structures on the nanoscale. However, current structure reconstruction and segmentation approaches suffer from either exclusion of small structures or strong dependence on a-priori knowledge. We propose reconstruction methods based on compressed sensing (CS) denoising in combination with the isodata threshold for segmentation. The methods are verified on artificial test data. For the denoising, a Haar dictionary and a KSVD dictionary learning on artificial data are used. Both methods perform significantly better than the reference algorithm, a linear density filter, in terms of root-mean-square deviation from the ground truth. Furthermore, exemplary results on real LM data of irradiated cell nuclei with Heterochromatin labeling make small structures visible that are suppressed by the reference method. CS denoising demonstrates promising results for reconstruction of LM data.

Big Data applications in science are producing huge amounts of data, which require advanced processing, handling, and analysis capabilities. For the organization of large scale data sets it is essential to annotate these with metadata, index them, and make them easily findable. In this paper we investigate two scientific use cases from biology and photon science, which entail complex situations in regard to data volume, data rates and analysis requirements. The LSDMA project provides an ideal context for this research, combining both innovative R&D on the processing, handling, and analysis level and a wide range of research communities in need of scalable solutions. To facilitate the advancement of data life cycles we present preferred metadata management strategies. In biology the Open Microscopy Environment (OME) and in photon science NeXus/ICAT are presented. We show that these are well suited for the respective data life cycles. To facilitate searching across communities we discuss solutions involving the Open Archive Initiative - Protocol for Metadata Harvesting (OAI-PMH) and Apache Lucene/Solr.

High-quality image denoising requires taking into account the dependence of the noise distribution on the original image. The parameters of this dependence are often unknown and we propose a new method to estimate them here. Using an optimization procedure, we find a variance-stabilizing transformation, which transforms the input image into an image with signal-independent noise. Principal component analysis of blocks of the transformed image allows estimation of the variance of the signal-independent noise so that the parameters of the original noise model can be computed. The image blocks for processing are selected in such a way that they have low stochastic texture strength but preserve the noise distribution. The algorithm does not require the original image to have homogeneous areas and can accurately process images with regular textures. It has high computational efficiency and smaller maximum estimation error compared with the state of the art. Our experiments have also shown that denoising with the noise parameters estimated by this method leads to the same results as denoising with the true noise parameters.

The problem of blind noise level estimation arises in many image processing applications, such as denoising, compression, and segmentation. In this paper, we propose a new noise level estimation method on the basis of principal component analysis of image blocks. We show that the noise variance can be estimated as the smallest eigenvalue of the image block covariance matrix. Compared with 13 existing methods, the proposed approach shows a good compromise between speed and accuracy. It is at least 15 times faster than methods with similar accuracy, and it is at least two times more accurate than other methods. Our method does not assume the existence of homogeneous areas in the input image and, hence, can successfully process images containing only textures.

This paper presents a real-time capable graphics processing unit (GPU)-based ultrasound simulator suitable for medical education. The main focus of the simulator is to synthesize realistic looking ultrasound images in real-time including artifacts, which are essential for the interpretation of this data. The simulation is based on a convolution-enhanced ray-tracing approach and uses a deformable mesh model. Deformations of the mesh model are calculated using the PhysX engine. Our method advances the state of the art for real-time capable ultrasound simulators by following the path of the ultrasound pulse, which enables better simulation of ultrasound-specific artifacts. An evaluation of our proposed method in comparison with recent generative slicing-based strategies as well as real ultrasound images is performed. Hereby, a gelatin ultrasound phantom containing syringes filled with different media is scanned with a real transducer. The obtained images are then compared to images which are simulated using a slicing-based technique and our proposed method. The particular benefit of our method is the accurate simulation of ultrasound-specific artifacts, like range distortion, refraction and acoustic shadowing. Several test scenarios are evaluated regarding simulation time, to show the performance and the bottleneck of our method. While being computationally more intensive than slicing techniques, our simulator is able to produce high-quality images in real-time, tracing over 5000 rays through mesh models with more than 2 000 000 triangles of which up to 200 000 may be deformed each frame.

Hardware failures occur especially due to external influences, component aging, or faulty interoperability. By testing, faulty components can be localized, allowing for fault isolation or repair. The contract testing strategy from software specifies component interoperability conditions, and systematically creates correspondent tests ensuring the operability of the system. We adapt contract testing to hardware, providing component specification, and monitoring thereof. Contract specification has to be specialized with requirements on the physical environment and component input signals. Contract testing is then executed through the monitoring of the contract parameters. Furthermore, to reason about the external cause of errors, signal faults are categorized. As a case study, we present an communication system. For this system, a contract is defined, and circuits implemented to perform contract testing and fault categorization. Communication faults, related to hardware errors and to sporadic environment disturbance, are injected in the developed system. These faults are completely detected, but can be only partially categorized by the monitoring approach.

A high level of psychomotor abilities often is essential for successfully operating many technical systems and especially medical and rehabilitation devices. This paper introduces an approach to enable a technical system to automatically assess its user's level of psychomotor abilities, so that it can adapt its level of automation and provide the user with more or less assistance depending on the individual user profile. For this purpose, a study has been conducted during which the motor abilities of the participants have been assessed and their wheelchair control behavior recorded. Bayesian networks (BN) and structural equation models (SEM) have been applied to model the relationships between the wheelchair control behavior and the motor abilities of the participants. The BN demonstrate usefulness and magnificent advantages compared to the SEM for modeling uncertainty in structure and parameter dependencies, which are shown by validation experiments. Although only a small amount of data samples (23 participants) was available for model generation, a target variable reflecting the user's precision ability was successfully classified based on real data input in more than 82% of cases.

Separating bone, calcification, and vessels in computer tomography angiography (CTA) allows for a detailed diagnosis of vessel stenosis. This paper presents a new, graph-based technique that solves this difficult problem with high accuracy. The approach requires one native data set and one that is contrast enhanced. On each data set, an attributed level-graph is derived and both graphs are matched by dynamic programming to differentiate between bone, on one hand side, and vessel/calcification on the other hand side. Lumen and calcified regions are then separated by a profile technique. Evaluation is based on data from vessels of pelvis and lower extremities of elderly patients. Due to substantial calcification and motion of patients between and during the acquisitions, the underlying approach is tested on a class of difficult cases. Analysis requires 3-5 min on a Pentium IV 3 GHz for a 700 MByte data set. Among 37 patients, our approach correctly identifies all three components in 80% of cases correctly compared to visual control. Critical inconsistencies with visual inspection were found in 6% of all cases; 70% of these inconsistencies are due to small vessels that have 1) a diameter near the resolution of the CT and 2) are passing next to bony structures. All other remaining deviations are found in an incorrect handling of the iliac artery since the slice thickness is near the diameter of this vessel and since the orientation is not in cranio-caudal direction. Increasing resolution is thus expected to solve many the aforementioned difficulties.

In this paper, we present a new method to model individual cognitive abilities, which were originally described by structural equation models and which can be used to reduce the possibility of human error when interacting with complex technical system. Our proposed model is based on hierarchical Bayesian networks and deals therefore with uncertain and noisy data. The validity of the method is demonstrated on the basis of simulations of real-life data. The inference from the Bayesian network demonstrates that estimating the user's cognitive abilities is possible in a reliable manner in real-time. Structural and data dependency analyses further show that less evidence data are necessary to provide cognitive profile classification. This especially highlights the advantage of our new method in comparison to the - in psychology - traditionally used structural equation models. Such cognitive profiling will allow adapting a system's demand character to the cognitive ability level of an individual user, which will make operating that system less error-prone, as many human errors result out of systems requesting more cognitive abilities than the operator has available.

We develop the first approach Tor interactive volume visualization based on a sophisticated rendering method of shear-warp type, wavelet data encoding techniques, and a trivariate spline model, which has been introduced recently. As a first step of our algorithm, we apply standard wavelet expansions to represent and decimate the given gridded three-dimensional data. Based on this data encoding, we give a sophisticated version of the shear-warp based volume rendering method. Our new algorithm visits each voxel only once taking advantage of the particular data organization of octrees. In addition, the hierarchies of the data guide the local (re)construction of the quadratic super-spline models, which we apply as a pure visualization tool. The low total degree of the polynomial pieces allows to numerically approximate the volume rendering integral efficiently. Since the coefficients of the splines are almost immediately available from the given data, Bernstein-Bezier techniques can be fully employed in our algorithms. In this way, we demonstrate that these models can be successfully applied to full volume rendering of hierarchically organized data. Our computational results show that (even when hierarchical approximations are used) the new approach leads to almost artifact-free visualizations of high quality for complicated and noise-contaminated volume data sets, while the computational effort is considerable low, i.e. our current implementation yields 1-2 frames per second for parallel perspective rendering a 2563 volume data set (using simple opacity transfer functions) in a 5122 view-port.

In this paper we present a method that allows removing bony structures and calcifications from CT angiography (CTA) data. It relies on two data sets, one with (CA) and one without (NCA) contrast agent. In contrast to current approaches, the algorithm uses local rigid matching of each single object. Therefore, it allows compensating motion of the patient during acquisition. Furthermore, the method uses a highly accurate surface-matching procedure so that local mismatches can be as small as 3 voxels and therefore even arteries near the bone are mostly separated correctly. The overall system requires about 20 minutes on a Pentium 4 2.4 GHz with 1 GByte RAM for a 700 MByte data set (CA+NCA) and is currently used in the clinic for evaluation purposes.

3D interactive technologies become more and more applicable in nowadays medicine. Fast 3D reconstruction of CT-data is one of the examples. We present an acceleration of relevant reconstruction algorithms using FPGAs. For this purpose we used a board based on Xilinx Virtex II6000. Our system is able to reconstruct a 2563 voxel volume from 120 projections (5122j in 9s, using the filtered backprojection algorithm. In addition, it is shown how such a system is used to accelerate iterative algorithms, whereby the forward projection is performed by graphics hardware. The performance for iterative approaches is 4 min for SIRT (10 iterations) and 3 minutes for OS-EM (6 iterations) where the main limitation is the forward projection.

Abstract A collection of techniques is developed in this paper for ray casting free‐form deformed‐volume objects with high quality and efficiency. The known inverse ray deformation approach is combined with free‐form deformation to bend the rays to the opposite direction of the deformation, producing an image of the deformed volume without generating a really deformed intermediate volume. The local curvature is estimated and used for the adaptive selection of the length of polyline segments, which approximate the inversely deformed ray trajectories; thus longer polyline segments can be automatically selected in regions with small curvature, reducing deformation calculation without losing the spatial continuity of the simulated deformation. We developed an efficient method for the estimation of the local deformation function. The Jacobian of the local deformation function is used for adjustments of the opacity values and normal vectors computed from the original volume, guaranteeing that the deformed spatial structures are correctly rendered. The popular ray casting acceleration techniques, like early ray termination and space leaping, are incorporated into the deformation procedure, providing a speed‐up factor of 2.34–6.56 compared to the non‐optimized case. Copyright © 2003 John Wiley & Sons, Ltd.

The aim of this project is to develop a new approach for the correction of bone deformities using a computer-aided surgery tool. We use measurement-based planning of corrections on CT data and an ultrasound-based navigation system for the real-time support of surgical operations. We implement a single-cut, close-and-open wedge osteotomy correction of complex 3D deformities. In the discussed method, virtual reality and image processing techniques allow planning and supporting of the operation on a personal computer. The designed software prototype significantly simplifies the process of osteotomy planning and allows surgeon training and assistance. The program reduces the total time of the operation planning and increases the accuracy of the correction significantly.

Abstract Volume rendering is a key technique in scientific visualization that lends itself to significant exploitable parallelism. The high computational demands of real‐time volume rendering and continued technological advances in the area of VLSl give impetus to the development of special‐purpose volume rendering architectures. This paper presents and characterizes three recently developed volume rendering engines which are based on the ray‐casting method. A taxonomy of the algorithmic variants of ray‐casting and details of each ray‐casting architecture are discussed. The paper then compares the machinefeatures and provides an outlook onfuture developments in the area of volume rendering hardware.