Computed tomography (CT) combined with deep learning (DL) has recently shown great potential in biomedical imaging. Complex DL models with varying architectures inspired by the human brain are improving imaging software and aiding diagnosis. However, the accuracy of these DL models heavily relies on the datasets used for training, which often contain low-quality CT images from low-dose CT (LDCT) scans. Moreover, in contrast to the neural architecture of the human brain, DL models today are dense and complex, resulting in a significant computational footprint. Therefore, in this work, we propose sparse optimizations to minimize the complexity of the DL models and leverage architecture-aware optimization to reduce the total training time of these DL models. To that end, we leverage a DL model called DenseNet and Deconvolution Network (DDNet). The model enhances LDCT chest images into high-quality (HQ) ones but requires many hours to train. To further improve the quality of final HQ images, we first modified DDNet's architecture with a more robust multi-level VGG (ML-VGG) loss function to achieve state-of-the-art CT image enhancement. However, improving the loss function results in increased computational cost. Hence, we introduce sparse optimizations to reduce the complexity of the improved DL model and then propose architecture-aware optimizations to efficiently utilize the underlying computing hardware to reduce the overall training time. Finally, we evaluate our techniques for performance and accuracy using state-of-the-art hardware resources. / Master of Science / Deep learning-based (DL) techniques that leverage computed tomography (CT) are becoming omnipresent in diagnosing diseases and abnormalities associated with different parts of the human body. However, their diagnostic accuracy is directly proportional to the quality of the CT images used in training the DL models, which is majorly governed by the radiation dose of the X-ray in the CT scanner. To improve the quality of low-dose CT (LDCT) images, DL-based techniques show promising improvements. However, these techniques require substantial computational resources and time to train the DL models. Therefore, in this work, we incorporate algorithmic techniques inspired by sparse neural architecture of the human brain to reduce the complexity of such DL models. To that end, we leverage a DL model called DenseNet and Deconvolution Network (DDNet) that enhances the quality of CT images generated by low X-ray dosage into high-quality CT images. However, due to its architecture, it takes hours to train DDNet on state-of-the-art hardware resources. Hence, in this work, we propose techniques that efficiently utilize the hardware resources and reduce the time required to train DDNet. We evaluate the efficacy of our techniques on modern supercomputers in terms of speed and accuracy.

This thesis delves into the advancement of 5G portable test-nodes in wireless communication systems with cognitive radio capabilities, specifically addressing the critical need for dynamic spectrum sensing and awareness at the radio receiver through AI-driven automatic modulation classification. Our methodology is centered around the transformer encoder architecture incorporating a multi-head self-attention mechanism. We train our architecture extensively across a diverse range of signal-to-noise ratios (SNRs) from the RadioML 2018.01A dataset. We introduce a novel transformer-based architecture with a gated mechanism, designed as a runtime re-configurable automatic modulation classification framework, which demonstrates enhanced performance with low SNR RF signals during evaluation, an area where conventional methods have shown limitations, as corroborated by existing research. Our innovative single-model framework employs distinct weight sets, activated by varying SNR levels, to enable a gating mechanism for more accurate modulation classification. This advancement in automatic modulation classification marks a crucial step toward the evolution of smarter communication systems. / Master of Science / This thesis delves into the advancement of wireless communication systems, particularly in developing portable devices capable of effectively detecting and analyzing radio signals with cognitive radio capabilities. Central to our research is leveraging artificial intelligence (AI) for automatic modulation classification, a method to identify signal modulation types. We utilize a transformer-based AI model trained on the RadioML 2018.01A dataset. Our training approach is particularly effective when evaluating low-quality signals using a gating mechanism based on signal-to-noise ratios, an area previously considered challenging in existing research. This work marks a significant advancement in creating more intelligent and responsive wireless communication systems.

As the field of mobile robotics rapidly expands, precise understanding of a robot’s position and orientation becomes critical for autonomous navigation and efficient task performance. In this thesis, we present a snapshot-based global localization machine learning model for a mobile robot, the e-puck, in a simulated environment. Our model uses multimodal data to predict both position and orientation using the robot’s on-board cameras and LiDAR sensor. In an effort to minimize localization error, we explore different sensor configurations by varying the number of cameras and LiDAR layers used. Additionally, we investigate the performance benefits of different multimodal fusion strategies while leveraging the EfficientNet CNN architecture as our model’s foundation. Data collection and testing is conducted using Webots simulation software, and our results show that, when tested in a 12m x 12m simulated apartment environment, our model is able to achieve positional accuracy within 0.2m for each of the x and y coordinates and orientation accuracy within 2°, all without the need for sequential data history. Our results demonstrate the potential for accurate global localization of mobile robots in simulated environments without the need for existing maps or temporal data.

In this work, we examine semi-supervised learning for Gait recognition with a limited number of labeled samples. Our research focus on two distinct aspects for limited labels, 1)closed-set: with limited labeled samples per individual, and 2) open-set: with limited labeled individuals. We find open-set poses greater challenge compared to closed-set thus, having more labeled ids is important for performance than having more labeled samples per id. Moreover, obtaining labeled samples for a large number of individuals is usually more challenging, therefore limited id setup (closed-setup) is more important to study where most of the training samples belong to unknown ids. We further analyze that existing semi-supervised learning approaches are not well suited for scenario where unlabeled samples belong to novel ids. We propose a simple prototypical self-training approach to solve this problem, where, we integrate semi-supervised learning for closed set setting with self-training which can effectively utilize unlabeled samples from unknown ids. To further alleviate the challenges of limited labeled samples, we explore the role of synthetic data where we utilize diffusion model to generate samples from both known and unknown ids. We perform our experiments on two different Gait recognition benchmarks, CASIA-B and OUMVLP, and provide a comprehensive evaluation of the proposed method. The proposed approach is effective and generalizable for both closed and open-set settings. With merely 20% of labeled samples, we were able to achieve performance competitive to supervised methods utilizing 100% labeled samples while outperforming existing semi-supervised methods.

In this thesis we aim to supplement the Climate and Economic Justice Screening Tool (CE JST), which assists federal agencies in identifying disadvantaged census tracts, by extracting five environmental features from Google Street View (GSV) images. The five environmental features are garbage bags, greenery, and three distinct road damage types (longitudinal, transverse, and alligator cracks), which were identified using image classification, object detection, and image segmentation. We evaluate three cities using this developed feature space in order to distinguish between disadvantaged and non-disadvantaged census tracts. The results of the analysis reveal the significance of the feature space and demonstrate the time efficiency, detail, and cost-effectiveness of the proposed methodology. / Master of Science / In this thesis we aim to supplement the Climate and Economic Justice Screening Tool (CE JST), which assists federal agencies in identifying disadvantaged census tracts, by extracting five environmental features from Google Street View (GSV) images. The five environmental features are garbage bags, greenery, and three distinct road damage types (longitudinal, transverse, and alligator cracks), which were identified using image classification, object detection, and image segmentation. We evaluate three cities using this developed feature space in order to distinguish between disadvantaged and non-disadvantaged census tracts. The results of the analysis reveal the significance of the feature space and demonstrate the time efficiency, detail, and cost-effectiveness of the proposed methodology.

This thesis investigates the performance of deep learning models, specifically Resnet50 and TransUnet, in semantic image segmentation on microwell images containing tumor and natural killer (NK) cells. The main goal is to examine the effect of only using bright-field data (1-channel) as input instead of both fluorescent and brightfield data (4-channel); this is interesting since fluorescent imaging can cause damage to the cells being analyzed. The network performance is measured by Intersection over Union (IoU), the networks were trained and using manually annotated data from Onfelt Lab. TransUnet consistently outperformed the Resnet50 for both the 4-channel and 1-channel data. Moreover, the 4-channel input generally resulted in a better IoU compared to using only the bright-field channel. Furthermore, a significant decline in performance is observed when the networks are tested on the control data. For the control data, the overall IoU for the best performing 4-channel model dropped from 86.2\% to 73.9\%. The best performing 1-channel model dropped from 83.8\% to 70.8\% overall IoU.

The Advanced Scatterometer (ASCAT) is a satellite-based remote sensing instrument designed for measuring wind speed and direction over the Earth's oceans. This thesis aims to expand and improve the capabilities of ASCAT by adding rain detection and advancing wind retrieval. Additionally, this expansion to ASCAT serves as evidence of Artificial Intelligence (AI) techniques learning both novel and traditional methods in remote sensing. I apply semantic segmentation to ASCAT measurements to detect rain over the oceans, enhancing capabilities to monitor global precipitation. I use two common neural network architectures and train them on measurements from the Tropical Rainfall Measuring Mission (TRMM) collocated with ASCAT measurements. I apply the same semantic segmentation techniques on wind retrieval in order to create a machine learning model that acts as an inverse Geophysical Model Function (GMF). I use three common neural network architectures and train the models on ASCAT data collocated with European Centre for Medium-Range Weather Forecasts (ECMWF) wind vector data. I successfully increase the capabilities of the ASCAT satellite to detect rainfall in Earth's oceans, with the ability to retrieve wind vectors without a GMF or Maximum Likelihood Estimation (MLE).

Numerous of high-impact applications involve predictive modeling of real-world data. This spans from hospital readmission prediction for enhanced patient care up to event detection in power systems for grid stabilization. Developing performant machine learning models necessitates extensive high-quality training data, ample labeled samples, and training and testing datasets derived from identical distributions. Though, such methodologies may be impractical in applications where obtaining labeled data is expensive or challenging, the quality of data is low, or when challenged with covariate or concept shifts. Our emphasis was on devising transfer learning methods to address the inherent challenges across two distinct applications.We delved into a notably challenging transfer learning application that revolves around predicting hospital readmission risks using electronic health record (EHR) data to identify patients who may benefit from extra care. Readmission models based on EHR data can be compromised by quality variations due to manual data input methods. Utilizing high-quality EHR data from a different hospital system to enhance prediction on a target hospital using traditional approaches might bias the dataset if distributions of the source and target data are different. To address this, we introduce an Early Readmission Risk Temporal Deep Adaptation Network, ERR-TDAN, for cross-domain knowledge transfer. A model developed using target data from an urban academic hospital was enhanced by transferring knowledge from high-quality source data. Given the success of our method in learning from data sourced from multiple hospital systems with different distributions, we further addressed the challenge and infeasibility of developing hospital-specific readmission risk prediction models using data from individual hospital systems. Herein, based on an extension of the previous method, we introduce an Early Readmission Risk Domain Generalization Network, ERR-DGN. It is adept at generalizing across multiple EHR data sources and seamlessly adapting to previously unseen test domains. In another challenging application, we addressed event detection in electrical grids where dependencies are spatiotemporal, highly non-linear, and non-linear systems using high-volume field-recorded data from multiple Phasor Measurement Units (PMUs). Existing historical event logs created manually do not correlate well with the corresponding PMU measurements due to scarce and temporally imprecise labels. Extending event logs to a more complete set of labeled events is very costly and often infeasible to obtain. We focused on utilizing a transfer learning method tailored for event detection from PMU data to reduce the need for additional manual labeling. To demonstrate the feasibility, we tested our approach on large datasets collected from the Western and Eastern Interconnections of the U.S.A. by reusing a small number of carefully selected labeled PMU data from a power system to detect events from another. Experimental findings suggest that the proposed knowledge transfer methods for healthcare and power system applications have the potential to effectively address the identified challenges and limitations. Evaluation of the proposed readmission models show that readmission risk predictions can be enhanced when leveraging higher-quality EHR data from a different site, and when trained on data from multiple sites and subsequently applied to a novel hospital site. Moreover, labels scarcity in power systems can be addressed by a transfer learning method in conjunction with a semi-supervised algorithm that is capable of detecting events based on minimal labeled instances. / Computer and Information Science

Ces travaux de recherche visent à développer des méthodes d’apprentissage automatique pour l’analyse de l’électroencéphalogramme (EEG) continu. L’EEG continu est une modalité avantageuse pour l’évaluation fonctionnelle des états cérébraux en réanimation ou pour d’autres applications. Cependant son utilisation aujourd’hui demeure plus restreinte qu’elle ne pourrait l’être, car dans la plupart des cas l’interprétation est effectuée visuellement par des spécialistes.Les sous-parties de ce travail s’articulent autour de l’évaluation pronostique du coma post-anoxique, choisie comme application pilote. Un petit nombre d’enregistrement longue durée a été réalisé, et des enregistrements existants ont été récupérés au CHU Grenoble.Nous commençons par valider l’efficacité des réseaux de neurones profonds pour l’analyse EEG d’échantillons bruts. Nous choisissons à cet effet de travailler sur la classification de stades de sommeil. Nous utilisons un réseau de neurones convolutionnel adapté pour l’EEG que nous entrainons et évaluons sur le jeu de données SHHS (Sleep Heart Health Study). Cela constitue le premier system neuronal à cette échelle (5000 patients) pour l’analyse du sommeil. Les performances de classification atteignent ou dépassent l’état de l’art.En utilisation réelle, pour la plupart des applications cliniques le défi principal est le manque d’annotations adéquates sur les patterns EEG ou sur de court segments de données (et la difficulté d’en établir). Les annotations disponibles sont généralement haut niveau (par exemple, le devenir clinique) est sont donc peu nombreuses. Nous recherchons comment apprendre des représentations compactes de séquences EEG de façon non-supervisée/semi-supervisée. Le domaine de l’apprentissage non supervisé est encore jeune. Pour se comparer aux travaux existants nous commençons avec des données de type image, et investiguons l’utilisation de réseaux adversaires génératifs (GANs) pour l’apprentissage adversaire non-supervisé de représentations. La qualité et la stabilité de différentes variantes sont évaluées. Nous appliquons ensuite un GAN de Wasserstein avec pénalité sur les gradients à la génération de séquences EEG. Le système, entrainé sur des séquences mono-piste de patients en coma post anoxique, est capable de générer des séquences réalistes. Nous développons et discutons aussi des idées originales pour l’apprentissage de représentations en alignant des distributions dans l’espace de sortie du réseau représentatif.Pour finir, les signaux EEG multipistes ont des spécificités qu’il est souhaitable de prendre en compte dans les architectures de caractérisation. Chaque échantillon d’EEG est un mélange instantané des activités d’un certain nombre de sources. Partant de ce constat nous proposons un système d’analyse composé d’un sous-système d’analyse spatiale suivi d’un sous-système d’analyse temporelle. Le sous-système d’analyse spatiale est une extension de méthodes de séparation de sources construite à l’aide de couches neuronales avec des poids adaptatifs pour la recombinaison des pistes, c’est à dire que ces poids ne sont pas appris mais dépendent de caractéristiques du signal d’entrée. Nous montrons que cette architecture peut apprendre à réaliser une analyse en composantes indépendantes, si elle est entrainée sur une mesure de non-gaussianité. Pour l’analyse temporelle, des réseaux convolutionnels classiques utilisés séparément sur les pistes recombinées peuvent être utilisés. / The objective of this research is to explore and develop machine learning methods for the analysis of continuous electroencephalogram (EEG). Continuous EEG is an interesting modality for functional evaluation of cerebral state in the intensive care unit and beyond. Today its clinical use remains more limited that it could be because interpretation is still mostly performed visually by trained experts. In this work we develop automated analysis tools based on deep neural models.The subparts of this work hinge around post-anoxic coma prognostication, chosen as pilot application. A small number of long-duration records were performed and available existing data was gathered from CHU Grenoble. Different components of a semi-supervised architecture that addresses the application are imagined, developed, and validated on surrogate tasks.First, we validate the effectiveness of deep neural networks for EEG analysis from raw samples. For this we choose the supervised task of sleep stage classification from single-channel EEG. We use a convolutional neural network adapted for EEG and we train and evaluate the system on the SHHS (Sleep Heart Health Study) dataset. This constitutes the first neural sleep scoring system at this scale (5000 patients). Classification performance reaches or surpasses the state of the art.In real use for most clinical applications, the main challenge is the lack of (and difficulty of establishing) suitable annotations on patterns or short EEG segments. Available annotations are high-level (for example, clinical outcome) and therefore they are few. We search how to learn compact EEG representations in an unsupervised/semi-supervised manner. The field of unsupervised learning using deep neural networks is still young. To compare to existing work we start with image data and investigate the use of generative adversarial networks (GANs) for unsupervised adversarial representation learning. The quality and stability of different variants are evaluated. We then apply Gradient-penalized Wasserstein GANs on EEG sequences generation. The system is trained on single channel sequences from post-anoxic coma patients and is able to generate realistic synthetic sequences. We also explore and discuss original ideas for learning representations through matching distributions in the output space of representative networks.Finally, multichannel EEG signals have specificities that should be accounted for in characterization architectures. Each EEG sample is an instantaneous mixture of the activities of a number of sources. Based on this statement we propose an analysis system made of a spatial analysis subsystem followed by a temporal analysis subsystem. The spatial analysis subsystem is an extension of source separation methods built with a neural architecture with adaptive recombination weights, i.e. weights that are not learned but depend on features of the input. We show that this architecture learns to perform Independent Component Analysis if it is trained on a measure of non-gaussianity. For temporal analysis, standard (shared) convolutional neural networks applied on separate recomposed channels can be used.

Fluorescence microscopy is an indispensable tool for biology to study the spatio-temporal dynamics of cells, tissues, and developing organisms. Modern imaging modalities, such as light-sheet microscopy, are able to acquire large three- dimensional volumes with high spatio-temporal resolution for many hours or days, thereby routinely generating Terabytes of image data in a single experiment. The quality of these images, however, is limited by the optics of the microscope, the signal-to-noise ratio of acquisitions, the photo-toxic effects of illumination, and the distortion of light by the sample. Additionally, the serial operation mode of most microscopy experiments, where large data sets are first acquired and only afterwards inspected and analyzed, excludes the possibility to optimize image quality during acquisition by automatically adapting the microscope parameters. These limits make certain observations difficult or impossible, forcing trade-offs between imaging speed, spatial resolution, light exposure, and imaging depth. This thesis is concerned with addressing several of these challenges with computational methods. First, I present methods for visualizing and processing the volumetric data from a microscope in real-time, i.e. at the acquisition rate of typical experiments, which is a prerequisite for the development of adaptive microscopes. I propose a low-discrepancy sampling strategy that enables the seamless display of large data sets during acquisition, investigate real-time compatible denoising, convolution, and deconvolution methods, and introduce a low-rank decomposition strategy for common deblurring tasks. Secondly, I propose a computational tractable method to simulate the interaction of light with realistically large biological tissues by combining a GPU-accelerated beam propagation method with a novel multiplexing scheme. I demonstrate that this approach enables to rigorously simulate the wave-optical image formation in light-sheet microscopes, to numerically investigate correlative effects in scattering tissues, and to elucidate the optical properties of the inverted mouse retina. Finally, I propose a data-driven restoration approach for fluorescence microscopy images based on convolutional neural networks (Care) that leverages sample and imaging specific prior knowledge. By demonstrating the superiority of this approach when compared to classical methods on a variety of problems, ranging from restoration of high quality images from low signal-to-noise-ratio acquisitions, to projection of noisy developing surface, isotropic recovery from anisotropic volumes, and to the recovery of diffraction-limited structures from widefield images alone, I show that Care is a flexible and general method to solve fundamental restoration problems in fluorescence microscopy.