The Shack-Hartmann wavefront sensor is a widely-used device to measure the light wavefront. Cur- rently, the sensor is used as a gradient sensor, which is achieved by placing the microlens array in the plane conjugate to the deformable mirror and the aberration. The resulting spot pattern is then transformed into a gradient through the use of centroiding: the measurement of the relative movement of all spots compared to the reference image. As an alternative to the traditional alignment, alternative alignment moves the Shack-Hartmann wave- front sensor: the CCD is now optically conjugated to the deformable mirror and the aberrations, and not the gradients but the curvature of the wavefront is measured. This changes the behaviour of the Shack-Hartmann sensor: instead of the amount of spots remaining the same and the CCD region oc- cupied by them changing, the amount of spots starts to change but the CCD region occupied by them remains the same in the face of aberrations. Curvature measurements can be used directly to actuate bimorph and membrane deformable mirrors, bypassing the need for complex control schemes. As an alternative to centroiding Fourier demodulation can also be used to convert the information measured by the wavefront sensor into gradients. For large aberrations where the spots stray out of their subapertures and the centroiding algorithm breaks down this method is especially suited. In this thesis it is shown that a synergy exists between alternative alignment and Fourier demodulation. Large aberrations no longer occupy a huge region on the CCD, instead changing the amount of spots, and the movement of the spots outside of their subapertures is no longer a problem as there is no need for rigidly defined subapertures. The combined use of alternative alignment and Fourier demodulation was tested on simulations, both for square and circular apertures, and for cases with extra added Poisson noise. Finally real-life experi- ments were also conducted using circular apertures.

Linear Parameter-Varying (LPV) systems can be used as a bridge to extend the well studied model based control methods of Linear Time-Invariant systems to certain nonlinear systems. Despite significant attention in literature over the last two decades, finding an efficient global state space identification algorithm remains an open problem. Furthermore, a common assumption has been that the scheduling signals governing the time-varying dynamics are known or measured exactly. These drawbacks are found to inhibit the number of use cases for model based LPV control. This thesis explores new ways of identifying LPV systems for more general nonlinear systems with limited information on the optimal scheduling variable and scheduling dimension.

To this end, two novel rank constrained least squares problems are presented to identify system matrices and scheduling signals from input and full state measurements, and a dictionary of possible scheduling signals. The use of the methods is demonstrated in a simulation experiment, and the results are compared to existing full state measurement methods.

At last, the full state measurement methods are coupled with a state sequence estimation method from literature in order to obtain a method for identification of quasi-LPV systems from Input-Output measurements. Here "quasi" indicates that the scheduling signal is dependent on the system state and input. The scheme is flexible, and a proof of concept is given on a small nonlinear system.

The study of tumor microenvironments (TMEs) and immune cell composition in cancer, a disease characterized by uncontrolled growth and spread of tumor cells, has become increasingly important for understanding tumor progression and patient outcomes. Tools such as the TME-Analyzer enable this kind of research, but their manual workflows highlight a common problem in medical imaging: the scarcity of labeled data. This limits the efficiency and applicability of supervised learning algorithms to improve such medical image analysis tools. Self-supervised learning algorithms offer a promising alternative by learning feature representations without requiring labeled data. This thesis aims to address the issue of label scarcity by exploring the potential of self-supervised learning models for TME analysis involving the classification of individual cells in multiplex immunofluorescence (MxIF) microscopy images of triple-negative breast cancer (TNBC) tissue.

To enable the learning of feature representations from MxIF images with an arbitrary number of color channels, this thesis proposes to pre-train an encoder network on every image channel separately according to the SimCLR algorithm and perform classification of multi-channel images by feeding the concatenated feature representation outputs of every channel to a classifier network — referred to as the Siamese configuration. A hyperparameter search is conducted to optimize the SimCLR encoder’s ability to learn high-quality feature representations of individual cells in MxIF images of TNBC tissue. Upon obtaining an optimal set of hyperparameters, the effectiveness of the learned feature representations in improving label-efficiency for individual cell classification is assessed.

The results demonstrate that the proposed Siamese configuration improves the accuracy of classifying the inflammation status of TNBC tumor sections by 2.63%. Additionally, the optimal set of hyperparameters identified through the search include the use of the normalized temperature cross-entropy loss function with low temperature and an added image intensity thresholding term, as well as zoom and brightness/contrast augmentations. Furthermore, the optimized self-supervised learning model improves label-efficiency for individual cell classification, maintaining performance with only 40% of labeled data, while performance drops only when the label percentage is reduced below this threshold.

Ultrafast ultrasound localization microscopy (ULM) is a super-resolved vascular imaging method that provides a 10-fold improvement in resolution compared to ultrafast ultrasound Doppler imaging. Because typical ULM acquisitions accumulate large numbers of synthetic microbubble (MB) tracks over hundreds of cardiac cycles, transient hemodynamic variations such as pulsatility get averaged out. Here we introduce two independent processing methods to retrieve pulsatile flow information from MB tracks sampled at kilohertz framerates and demonstrate their potential on a simulated dataset. Our first approach filters out ULM localization grid artifacts and successfully recovers the pulsatility fraction Pf with a root mean square error of 3.3%. Our second approach relies on the derivation of the velocity distribution of MBs as observed from a stationary observer. We show that pulsatile flow gives rise to a bimodal velocity distribution with peaks indicating the maximum and minimum velocity of the cardiac cycle. Measuring the locations of these peaks, we successfully estimated Pf with an error of 5.2%. Last, we evaluated the impact of the MB localization precision σ on our ability to retrieve the bimodal signature of a pulsatile flow. Together, our results demonstrate that pulsatility can be retrieved from high framerate ULM acquisitions and that the estimation of the pulsatility fraction improves with MB localization precision.

Meditation is a contemplative practice that is believed to entail attentional and emotional regulation. One of the biggest challenges in developing personalized, accessible healthcare options with meditation is finding understandable features that signify whether someone is meditating. Specifically, there is no consensus on a feature resulting from the electroencephalogram (EEG) in the current body of literature on meditation.

In this thesis, I propose a dynamic systems analysis on EEG data to obtain a dynamic feature capable of distinguishing meditation from an eyes-closed resting baseline. I gathered the EEG data at TNO (Dutch Organisation for Applied Scientific Research) from twenty-two participants during a sixteen-minute loving-kindness meditation and two two-minute baselines. The proposed methodology characterizes temporal and spatial characteristics of the EEG simultaneously by approximating the EEG dynamics with a linear model on short time windows. I assess changes among three features: the frequency and magnitude of the oscillatory dynamics and the corresponding active electrodes.

The analysis can identify changes in EEG dynamics for each individual. Across all participants, regions associated with vision and language processing were active throughout the experiment. Notably, attention-related regions were more involved during meditation than rest. Moreover, the results show a shift in active regions throughout the meditation and the baselines for several participants.

Moreover, the thesis investigates the sensitivity of the analytical approach to changes in the electrode subset used for the analysis. For each participant, I constructed a subset of electrodes that were most involved in the changing EEG dynamics. The personalized subset was most sensitive to changes between meditation and rest, compared to other subsets based on commercial wearable EEG headsets. Finally, I compare the findings of the dynamic systems approach to a conventional analytical approach, and the participants’ emotional ratings inquired in subjective questionnaires. Unfortunately, from the current data, there appears to be no relationship between the proposed features and the conventional measures or the subjective questionnaires.

Specular light reflections are the mirror-like reflections from a material interface. They appear in the observation of any illuminated surface. Specular reflections can be set apart from the diffuse reflection type, which has a random distribution of reflection directions. The radiance of the specular reflected light is governed by the Fresnel ratio and depends on both geometrical and spectral properties. A clear distinction can be made between the reflection properties of dielectric and the much more reflective non-dielectric material types. In certain cases the presence of specular reflections is useful, for example in object identification or computer graphics applications. However, often the presence of specular reflection causes bright spots on the image of an object, which results in loss of detail in these areas. Particularly in computer vision applications it is important to only observe the intrinsic diffuse reflection. Hence, a variety of specular reflection removal methods have been developed. Color-space analysis specular reflection removal methods are able to accurately recover the intrinsic diffuse color of the image of an object. Optimization based methods based on a non-negative matrix factorization of the data are considered. The reduced characteristic representation of the data allows the specular and diffuse reflection component to be separated more easily. A multi-image approach is considered because it provides more available information on the intrinsic diffuse reflection component, which is typically a relatively weak signal for non-dielectric surface reflections. A specular reflection removal method is developed, that combines a new reflection model for colored illumination with a sparse Non-negative Matrix Factorization (NMF) optimization in a multi-image framework. The proposed method is evaluated on synthetic data and real data images of non-dielectric materials acquired conform the proposed reflection model.

Inertial measurement units (IMUs) are getting more and more incorporated into our lives due to their improving accuracy, lower cost and smaller sizes. Applications for inertial-based orientation estimation can already be found in the field of computer vision, aerospace engineering, robotics, navigation, biomechanics, health monitoring and sports. In order to enable non-intrusive, long-term tracking, this study addresses the problem of orientation estimation using a loosely attached IMU. To this end, a hybrid model is proposed combining
classical orientation methods with machine learning (ML) techniques. Previous works have focused only on parts of this task, which is why to date, there are little to no solutions to this problem. In this study, a model is designed containing three steps that are shown to be beneficial for reducing the estimation error. With the first step, the model becomes robust to a misaligned mounting of the loosely attached IMU and it allows to identify which part of the error can be contributed to the average misalignment. For the second step, two
algorithms are presented which make the model invariant to the absolute heading direction of the recorded data. In the third step, the data is prepared such that the size of the search space is reduced by a factor 2. Therefore, less data is required for learning the problem or, with an equal amount of data, a bigger part of the process can be discovered. Besides these steps, it is shown that the use of physics-based knowledge in the form of orientation estimates,
as opposed to raw inertial data as input features for the ML method, is beneficial.

This MSc thesis presents the development of a viewpoint optimization framework to face the problem of detecting occluded fruits in autonomous harvesting. A Deep Reinforcement Learning (DRL) algorithm is developed in order to train a robotic manipulator to navigate to occlusion-free viewpoints of the tomato-target. Two Convolutional Neural Networks (CNN), You Only Look Once (YOLO) version 3 and Mask Regional Convolutional Neural Network (Mask R-CNN), are trained and evaluated in order to obtain visual information from the tomatoes. The two trained CNN achieve high detection accuracy and surpass other fruit detection methods. The instance segmentation Mask R-CNN result and an image processing algorithm are applied in an occlusion modelling method. The vision-based reward formulation for the DRL algorithm is closely related to the occlusion modelling metric. The DRL algorithm is trained and evaluated in a simulation environment by mounting a camera on a robotic arm. The use of different reward scheme and exploration strategy during DRL training shows their effect on training performance. The DRL training performance is assessed by illustrating the maximum visible tomato-target fraction per episode, the steps needed per episode, the trajectory of the robot's end-effector, and the initial and final tomato-target viewpoint in each episode. The evaluation results show a satisfactory performance, which depends on the reward and exploration strategy, as in the majority of the cases the robot can fully see the tomato-target after a few steps.

General anesthesia (GA) is an important medical procedure that induces unconsciousness to patients during surgery. Consciousness is a salient feature of the brain, whose neurophysiological features are difficult to be distinguished from unconsciousness. Though it can be defined as an event arising due to interactions in the nervous system, it entirely is not a reliable mechanism. Thus, tracking changes in the brain waves caused by GA is a challenging problem in neuroscience. The exact mechanism to quantify the state of the brain and to distinguish between conscious and unconscious brain is still difficult. Specific features to characterize the state of the brain from the patterns of the brain signal is challenging. Present-day depth of anesthesia monitors index values does not quantify the state of the brain.

An alternative approach is to use dynamical systems theory to assess the underlying dynamics of the brain with imaging technology (e.g., electroencephalographic and electrocorticographic data). Previous results from the literature suggest that stability can play a role in the characterization of unconsciousness. This thesis proposes a detailed study that focus on dynamical systems properties that go beyond stability. In particular, the proposed methodology aims to assess which regions of the brain intervene in the process of consciousness and unconsciousness, as well as quantify how often they interact with each other. Specifically, the approach seeks to leverage the eigenstructure of the underlying approximation of the neural activity captured from intracranial electrocorticographic data.

Our results show that it is possible to differentiate between anesthetic stages of the brain using eigendecomposition. This was possible through a framework that provides a regularised way to sparsify the state estimates of electrocorticographic (ECoG) signal to get a model for analysis of changes in the brain waves affected by GA. Later to look at the eigenvalues and eigenvectors, which gives the frequency of oscillation and direction between different regions of the brain, respectively. It was also observed that the pattern in the evolution of eigenvalues during different anesthetic stages could be able to interpret if the subject was under anesthesia or not.

The scientific analysis of historical paintings has been traditionally restricted to the analysis of paint cross-section samples. This invasive method provides extensive information but is inherently limited in scope due to the extreme heterogeneity of paintings. In the last decade, non-invasive spot analyses and spectral imaging methods have become increasingly widespread in cultural heritage science. Two of these methods are macroscopic X-ray fluorescence imaging spectroscopy (MA-XRF) and reflectance imaging spectroscopy (RIS). These methods allow for 2D-scanning the entire surface of a painting and provide complementary information on elemental and molecular composition and distribution of the paint. However, these methods are often used only for qualitative analysis of the paint based on relative distribution maps, revealing only limited information about the paint layer stratigraphy. This thesis is an exploration of a combined approach for quantitative analysis of paint composition and layer stratigraphy using MA-XRF and RIS. The research used a set of specially prepared paint samples of mixtures and multiple layer applications based on historically relevant pigments which were scanned using MA-XRF and RIS in the visible and Near IR range (400-2500 nm). The spectral data acquired were processed and analyzed in a variety of ways, including Non-negative Matrix Factorization, Non-Linear Least Square Fitting, among other methods; in an attempt to gather quantitative compositional and stratigraphic data. In these trials, using characteristic reflectance features in the visible range together with a comparison of highly and lightly absorbed X-ray fluorescence lines allowed for the identification and quantification of surface specific compounds related to the top paint layer. Further comparison of fluorescence lines and absorption features in the Near IR range provided a potential avenue for quantification of subsurface paint layers. The results confirm that the combination of these methods allows to reveal the paint stratigraphy. The project provides samples and data sets which may serve as the basis for the development of a robust algorithm to address this issue in the future.

One-Shot Neural Architecture Search (NAS) is a promising method to significantly reduce search time without any separate training. It can be treated as a Network Compression problem on the architecture parameters from an overparameterized network. However, there are two issues associated with most one-shot NAS methods. First, dependencies between a node and its predecessors and successors are often disregarded which result in improper treatment over zero operations. Second, architecture parameters pruning based on their magnitude is questionable. In this thesis, classic Bayesian learning approach is applied to alleviate these two issues. Unlike other NAS methods, we train the over-parameterized network for only one epoch before update network architecture. Impressively, this enabled us to find the optimal architecture in both proxy and proxyless tasks on CIFAR-10 within only 0.2 GPU days using a single GPU. As a byproduct, our approach can be transferred directly to convolutional neural networks compression by enforcing structural sparsity that is able to achieve extremely sparse networks without accuracy deterioration.