1,721,319 research outputs found
A parameter-optimizing model-based approach to the analysis of low-SNR image sequences for biological virus detection
Full text linkThis thesis presents the multi-objective parameter optimization of a novel image analysis process. The focus of application is automatic detection of nano-objects, for example biological viruses, in real-time. Nano-objects are detected by analyzing time series of images recorded with the PAMONO biosensor, after parameters have been optimized on synthetic data created by a signal model for PAMONO. PAMONO, which is short for Plasmon-Assisted Microscopy of Nano-Sized Objects, is a biosensor yielding indirect proofs for objects on the nanometer-scale by measuring the Surface Plasmon Resonance (SPR) effects they cause on the micrometer scale. It is an optical microscopy technique enabling the detection of biological viruses and other nano-objects within a portable device. The PAMONO biosensor produces time series of 2-D images on the order of 4000 half-megapixel images per experiment. A particular challenge for automatic analysis of this data emerges from its low Signal-to-Noise Ratio (SNR). Manual analysis takes approximately two days per experiment and analyzing person. With the automatic analysis process developed in this thesis, occurrences of nano-objects in PAMONO data can be counted and displayed in real-time while measurements are being taken. Analysis is divided into a GPU-based detector aiming at high sensitivity, complemented with a machine learning-based classifier aiming at high precision. The analysis process is embedded into a multi-objective optimization approach that automatically adapts algorithm choice and parameters to changes in physical sensor parameters. Such changes occur, for example, during sensor prototype development. In order to automatically evaluate the objectives undergoing optimization, a signal model for the PAMONO sensor is proposed, which serves to synthesize ground truth-annotated data. The parameters of the analysis process are optimized on this synthetic data, and the classifier is learned from it. Hence, the signal model must accurately mimic the data recorded by the sensor, which is achieved by incorporating real sensor data into synthesis. Both, optimized parameters and the learned classifier, achieve high quality results on the real sensor data to be analyzed: Nano-objects with diameters down to 100nm are detected reliably in PAMONO data. Note that the median SNR over all nano-objects to be detected was below two in the examined experiments with 100nm objects. While the presented analysis process can be used for real-time virus detection in PAMONO data, the optimization approach can serve in accelerating the advancement of the sensor prototype towards a final setup of its physical parameters: In this scenario, frequent changes in physical sensor parameters make the automatic adaptation of algorithmic process parameters a desirable goal. No expertise concerning the underlying algorithms is required in these use cases, enabling ready applicability in a lab scenario
Perceptual aspects of sound scattering in concert halls
No full textThis work aims to expand the understanding of sound scattering in architectural spaces and the comprehension of its influence on the auditory perception in concert halls. The notion of scattering coefficient, which numerically represents the physical phenomenon of sound scattering, constitutes the main paradigm for the entire work. In a first part, the scattering coefficient is presented in its meaning and implications, providing both the mathematical formulation and the empirical evaluation. Scattering coefficients of new objects, such as pieces of furniture, have been for the first time determined, hence the foundations for a new scattering coefficient open database is laid. A new solution for avoiding recurrent measurement inaccuracies is presented by means of an improved measurement setup, which consists of a revised scale model reverberation chamber. The benefit of having more accurate acoustic computer simulations by using a wider set of experimental data for scattering coefficient is proved by a case study of classroom acoustics. The implementation of scattering coefficient in different room acoustic computer software is shown and discussed by using a concert hall as a case study. In a second part, the relationship between scattering coefficient and auditory perception is explored. Binaural impulse responses have been determined for three different scenarios, such as two virtual enclosed spaces and one real concert hall, and convolved with music samples to be used in listening tests. Results from listening tests show how changes in scattering coefficient of diffusing surfaces affect the perception of music among the audience in concert halls. A difference limen for scattering coefficient is determined by means of auralized binaural room impulse responses, which have been obtained under different scattering conditions. Results from listening tests are shown and discussed
Signal enhancement and signal reconstruction for diffusion imaging using deep learning
No full textDiffusion imaging, which is based on magnetic resonance imaging (MRI), allows a reconstruction of neural pathways inside the human brain and can, therefore, be employed to investigate neuronal diseases. However, a major disadvantage of this method is that long acquisition times are required to achieve high accuracy. Another disadvantage is that MRI systems from different manufacturers differ considerably from each other, which is why the resulting signals and the resulting characteristics often cannot be compared across devices. This is particularly important for research of neural diseases conducted based on large group studies. This is why this dissertation presents and evaluates novel deep learning based approaches for solving the obstacles mentioned above. The first main part of this thesis presents a new deep learning method, which was developed especially for spherical signals (as they occur in diffusion imaging). This involves the inclusion of previously unused information through a convolution on the signal's sphere. Furthermore, various activation functions are evaluated concerning their applicability on a measured diffusion signal. The second part focuses on signal harmonization, i.e., the homogenization of different MRT systems. The method developed here uses a training database of traveling subjects who have been scanned on both scanners to learn a mapping of signals from the first to the second scanner. It is evaluated using various essential metrics that are often used in clinical practice. This allows a minimizing of differences between two MRI systems. The third part of this dissertation presents a new method which interpolates the measured signal based on a high-resolution training database. Since this interpolation is learned on a signal basis, no assumptions about the underlying microstructure or physical properties need to be made. Subsequently, this approach can be integrated into state-of-the-art reconstruction pipelines. As a result, the measurement time can be considerably reduced and the signal quality substantially improved. The last part of this thesis addresses the reconstruction of the diffusion orientations. Since different diffusion orientations might occur within a voxel, two methods were developed and evaluated that predict the fiber orientation distribution function based on training data. Both methods achieve a higher accuracy in comparison to the evaluated state-of-the-art method, especially if very complex signals occur or if only a few gradient directions have been acquired. In this thesis, new methods for signal reconstruction, signal interpolation and signal harmonization using deep learning are presented in the context of diffusion imaging. It is shown that an inclusion of additional information is of great importance and leads to a distinctly improved result, especially for signal interpolation but also in signal harmonization
Local feature description with invariance against affine projection
No full textWhile the understanding of the image content is relatively easy for most humans, an automatic analysis is challenging for a computer vision system. To describe the content of an image, local features, which can be understood as mathematical representations of image regions, are used frequently. The image description should be invariant against photo- and geometric image distortions which typically occur when the illumination or the viewing angle during image acquisition changes.This thesis focuses on increasing the robustness of local features under geometric, or more specific, affine image projection, in the context of object detection.To increase robustness of local features against geometric image distortions, the affine invariant coordinate transformation is developed. The affine invariant coordinate transformation is an iterative normalization algorithm which exploits local properties of an image region to normalize it, such that two image regions captured from different viewpoints are identical, up to a rotational transformation after normalization. It can be combined with various feature detection and feature extraction algorithms and used for both globally and locally distorted images.For the detection of objects with various affine projections in different image recordings, the correspondence consensus merging is developed. The correspondence consensus merging uses the normalization matrices of two corresponding features to estimate the projection between the objects. Based on the assumption that features belonging to one semantic object are projected similarly across the images, the correspondence consensus merging groups correspondences whose projection estimates are similar. The algorithm provides reliable object detection results even when the established correspondences between the two images contain plenty false correspondences and can also be used to distinguish between correct and false feature correspondences.The developed algorithms are evaluated on synthetically warped as well as on camera captured image pairs with global and local geometric projection and compared against state of the art methods for affine invariant feature extraction, respectively for feature grouping. It is shown, that especially for images with local geometric projections, the presented algorithm is superior to the state of the art. Furthermore, it is shown that the presented grouping of feature correspondences allows for reliable object detection
Perceptual aspects of sound scattering in concert halls
No full textThis work aims to expand the understanding of sound scattering in architectural spaces and the comprehension of its influence on the auditory perception in concert halls. The notion of scattering coefficient, which numerically represents the physical phenomenon of sound scattering, constitutes the main paradigm for the entire work. In a first part, the scattering coefficient is presented in its meaning and implications, providing both the mathematical formulation and the empirical evaluation. Scattering coefficients of new objects, such as pieces of furniture, have been for the first time determined, hence the foundations for a new scattering coefficient open database is laid. A new solution for avoiding recurrent measurement inaccuracies is presented by means of an improved measurement setup, which consists of a revised scale model reverberation chamber. The benefit of having more accurate acoustic computer simulations by using a wider set of experimental data for scattering coefficient is proved by a case study of classroom acoustics. The implementation of scattering coefficient in different room acoustic computer software is shown and discussed by using a concert hall as a case study. In a second part, the relationship between scattering coefficient and auditory perception is explored. Binaural impulse responses have been determined for three different scenarios, such as two virtual enclosed spaces and one real concert hall, and convolved with music samples to be used in listening tests. Results from listening tests show how changes in scattering coefficient of diffusing surfaces affect the perception of music among the audience in concert halls. A difference limen for scattering coefficient is determined by means of auralized binaural room impulse responses, which have been obtained under different scattering conditions. Results from listening tests are shown and discussed
Steigerung der Energieeffizienz des pneumatischen Fasertransportes
No full textThe textile industry is one of the most energy-intensive sectors of the manufacturing industry. In light of rising energy prices, energy consumption is an important investment criterion. About 45 – 50 % of the energy costs in fibre preparation are accounted for by pneumatic fibre transport. In order to avoid clogging of the conveying systems as well as damage and compaction of the fibres, the conveying fans are usually overdimensioned by about 30 % and are operated at maximum speed. The possibilities for geometric optimisation of the conveying systems are limited due to constructional boundaries. The greatest potential is to be achieved by reducing the air velocity and thus the speed of the conveying fan. The aim of the development is to reduce the air speed by at least 30 %. In the dissertation, a monitoring and control system is developed to control the speed of the conveying fan in an energy-optimised manner depending on the transport state of fibres and flakes. The selection of the sensors and the monitoring strategy is carried out according to the Nine-Step-Tool method, which is updated and extended in the thesis. The most significant characteristic variables for the transport condition are the fibre distribution factor and the temporal fluctuation of the fibre flow. A system model is derived from the tests on an industrial fibre conveying system. For the detection of the transport state, a Sugeno network is developed and trained with model data from the system model. The validation of the state monitoring is carried out using both the model data and the raw data from the industrial test. Furthermore, a two-degree-of-freedom control is developed and validated on the model. The reduction of the air velocity by 30 % leads to a reduction of the energy consumption by about 65 %. For a medium-sized conveying fan with a nominal energy input of 12.5 kW, this results in an annual savings potential of around 9,000 €. The payback period is 1 year. In industrial trials, energy savings of up to 74 % are possible, depending on the desired state of transport
Medical image segmentation: the potential of convoluational neural networks and transformers
No full textMedical image segmentation, a crucial aspect of computer-aided diagnosis, plays a vital role in accurately delineating anatomical structures and regions of interest (ROIs). However, conventional segmentation methods are time-consuming, labour-intensive, and error-prone, relying heavily on manual labelling and clinical expertise. The complexities of various imaging modalities add further challenges to the process. To overcome these limitations, researchers have turned their attention to deep learning models, specifically convolutional neural networks (CNNs) and Vision Transformers (ViTs). While CNNs have found widespread use in medical image segmentation, their consecutive convolution operations have limitations in modelling long-term dependencies and capturing global context. On the other hand, ViT offers an alternative architecture with properties such as long-range dependency modelling, weak inductive bias, and noise robustness, making them well-suited for medical image segmentation. However, ViTs may encounter difficulties in modelling local representations due to the tokenization process. This thesis delves into three unique pathways to overcome the constraints of CNNs and Transformers, all while attaining higher accuracy and model simplicity. First, it examines the significance of texture and shape representation as they provide valuable insights into tissue structures. A comparison between CNNs and ViTs is made concerning their ability to capture structural, texture, and shape information. Techniques such as shape-attention, style matching, and frequency recalibration or self-attention module redesign are explored to enhance their representation of these features. Secondly, the fusion of CNNs and Transformers is investigated to achieve precise and contextually aware segmentation by leveraging the advantages of each architecture. Finally, a comprehensive analysis is conducted on utilizing ViTs in their pure form, without incorporating CNN modules. Three novel approaches are proposed to achieve linear complexity and capture multi-scale representations for semantic segmentation. In conclusion, this thesis significantly contributes to the field of medical image segmentation by addressing various challenges and providing solutions in each direction. By exploring innovative methods and combining the strengths of CNNs and Transformers, this research represents a significant step forward in applying deep learning algorithms to clinical practice
Interpretable image features and stain-independent machine learning methods for automated analysis of renal histopathology
No full textHistological whole slide images are conventionally analyzed visually by clinical experts, which is a highly labor and time intensive procedure. However, analyzing the extensive information on such images in an automated way has been difficult so far due to the vast sizes of the images, tissue-, pathological-, staining-variations, and so on. These factors account for some of the most important limitations in histopathological analyses. In this work, the goal is to overcome some of these limitations by automating histopathological analyses with the following focus:1. To extract quantitative and interpretable features to expand prior biological knowledge, in an -omics like approach, “Pathomics”.2. To make automatic histopathological analyses stain-independent and exploit information contained in stains to improve automatic analyses. All methods in this work are based on renal images obtained from mice. For the first part, we developed a novel pipeline that extracts a comprehensive set of visual features as well as sub-visual features. Here, visual features are those which are detectable by a pathologist, and sub-visual features are those which are not discernible by human experts. A large set of features (intensity, textural, shape, morphological, color, and nuclei-related) were extracted from several renal compartments including glomerular tuft, Bowmann’s Capsule, tubule, interstitium, arterial blood vessels and their lumen. This approach, similar to Radiomics, is referred to as Pathomics in pathology. We defined feature selection methods to extract the most informative and discriminative features and performed statistical analyses to understand the relation of the extracted features, both individually, and in combinations, with tissue morphology and pathology. In the presented case-study, we highlight features that are affected in each compartment for experimental unilateral ureteral obstruction and their contralateral tissue for comparative analyses. In this way, prior biological knowledge is confirmed and presented in a quantitative way, alongside with novel findings. The proposed approach provides a quantitative, reproducible, and rater-independent characterization of whole slide images, e.g. for quantitatively assessing disease-specific changes in histopathology. To address the second goal, we developed Generative Adversarial Networks based methods which facilitate virtual stain translations. This makes it possible to perform stain independent analyses, which overcomes a major limitation in automatic histopathological analyses. In this work, we focus on utilizing the virtually stained images thus obtained to further improve the performance of deep learning algorithms. To this end, we introduced the idea of “image enrichment”: we merge virtually stained images with the original image to create a multi-channel image, which provides an “enriched” image with a higher information content. We prove the gain of information by showing that deep neural networks trained with the enriched images show higher segmentation accuracies
Efficient probability distribution function estimation for energy based image segmentation methods
No full textThis work investigates efficient energy based image segmentation methods when only little prior knowledge about an object is given. It is concentrated on implementing stochastic approaches, utilizing the channel framework for modelling probability distribution functions within the data terms of energy functionals. In addition, an approach is developed that uses the block Gibbs sampler within an edge-detection algorithm. Traditionally, in related segmentation methods, PDFs are approximated using either histograms or kernel density estimation (KDE). The channel framework tries to combine advantages of both, keeping the complexity similarly low as in the histogram approach, while tending to approach similar performance as KDE methods. The channel framework can be interpreted as a soft histogram where bins overlap in the amplitude space. This allows sub-bin accuracy and, in segmentation methods, less bins are necessary to approximate the PDF better than by original histograms. Another interpretation of the channel framework is that it is a discrete version of the KDE approach. This way, the channel framework gives reasonably fast performance even for large images and the speed is not image content dependent. On the other hand, to achieve similar performance as KDE, a large number of channels is required. It is shown that an efficient implementation is still possible by using a look-up table, which however requires an increased amount of memory as the cost for the decrease in computational complexity. The following variants of image segmentation methods have been extended to use the channel framework: automatic/unsupervised image segmentation in two regions using the Chan-Vese functional; supervised image segmentation in two regions using the Chan-Vese functional; interactive segmentation in multiple regions using the Potts model. All methods have been tested on different evaluation data sets. In most cases, the channel framework has given comparable results to the state of the art, while being more efficient to implement.In addition to segmentation methods, an edge-detection method is introduced, which provides a stochastic approach to solving the Ambrosio and Tortorelli functional by using the block Gibbs sampler. Unlike previous approaches, whole images are sampled at once and not pixel-wise. This leads to a faster converging solution than conventional convex optimization
Automated analysis of haematopoietic cells in bone marrow microscopy images
No full textTraditionally, haematologists diagnose haematopoietic diseases, such as leukaemia, based on the distribution of cell types in bone marrow microscopy images. The automation of this procedure is beneficial in multiple aspects: higher throughput, objective results, and a larger number of cells considered in the statistical analysis of the cell distribution. In this dissertation, dedicated solutions to a wide range of challenges offered by haematopoietic cell data are presented. First, limitations with state-of-the-art cell detection and classification approaches are addressed, for example through Circular Anchors, which considerably reduce the number of false positives during cell detection. This is achieved through better cell representations with circular instead of rectangular anchors. Furthermore, several techniques to incorporate weak annotations are proposed along with novel techniques for reliable cell detection using U-Net. Additionally, this dissertation proposes problem-specific techniques to solve various challenges of haematopoietic cell data. These challenges include visual variabilities caused by inhomogeneities in the staining process and the ordinal aspect of maturity stages of cells within one lineage. In addition, improvements to state-of-the-art methods for semi-supervised learning, which enable new approaches for network pre-training, are presented. Another significant contribution are the novel embedding learning techniques, which incorporate domain knowledge into neural network training. This dissertation introduces the idea of Embedding Guides, which encode different types of cells as points in a two-dimensional vector space. These are utilised by novel methods that enforce such guides in the training of suitable embedding spaces. Other novel methodologies proposed are Spatial Maturity Regression, a regularisation of the embedding space for the maturity in individual cell lineages, and Automatic Latent Interventions, a technique for generic improvements of the embedding space. The proposed methods show considerable improvements of raw classification accuracy and/or secondary scores, which are highly relevant for diagnoses. The approaches developed here suggest that a framework for the automated analysis of bone marrow microscopy images can be developed. This dissertation also includes a discussion on the suitability of automatic analyses in clinical practice as well as considerations for a tool designed for this purpose
- …
