Saudi Cultural Missions Theses & Dissertations
Permanent URI for this communityhttps://drepo.sdl.edu.sa/handle/20.500.14154/10
Browse
7 results
Search Results
Item Restricted Optimising Domain-Adversarial Neural Network (DANN) for Automated OCT-based Classification of Coronary Atherosclerotic Plaque Types Using Labelled Pig and Unlabelled Patient OCT Pullbacks(Queen Mary University of London, 2024-08) Alharthi, Hatem; Krams, RobThis research investigates the optimisation of a Domain-Adversarial Neural Network (DANN) for automated classification of five atherosclerotic plaque types using intracoronary Optical Coherence Tomography (OCT) pullbacks. Leveraging histologically co-registered labelled pig OCT images and unlabelled human OCT images obtained from Coronary Artery Disease (CAD) patients, the research focused on enhancing DANN’s ability to extract domain-invariant feature representations, thus adapt to domain shifts. A key innovative fine-tuning strategy was implemented using selective fine-tuning of the last four layers of a pre-trained DenseNet-121 model, which significantly improved the model's performance, achieving an average AUC-ROC of 0.935. The incorporation of a Gradient Reversal Layer (GRL) effectively mitigates domain discrepancies, as evidenced by a decrease in Proxy A-Distance from 2.0 to 0.66, and clearly visualised using t-SNE. The model demonstrated high testing sensitivity and specificity across all plaque types and specifically in identifying Thin-Cap Fibroatheroma (TCFA) plaque with 100% accuracy and sensitivity on our pig source data, indicating its potential for clinical application in cardiology. While the study acknowledges limitations such as dataset size and the empirical approach to model tuning, the findings contribute valuable insights into the role of domain adaptation in medical imaging, offering a robust framework for future research and clinical implementation.18 0Item Restricted Forecasting Solar Power Time Series: Strategies For Multi-Modal Data Fusion, Feature Relevance, and Sparse Data Management(RMIT University, 2024-06) Almaghrabi, Sarah; Hamilton, Margaret; Saiedur Rahaman, Mohammad; Rana, MashudThe forecasting of solar photovoltaic power (SPVP) is a significant challenge. Solar is the least reliable renewable energy, as it depends on the weather, among other things. However, it is also one of the cheapest sources if it can be harnessed, particularly during daylight hours when people work and use electricity. The ultimate aim of forecasting solar power using deep learning (DL) techniques is to enable the aggregated use of solar power stations by day, supplemented by alternative sources of electricity whenever solar energy is forecast to fall below a particular level. The more accurate the solar power predictions, the better the use and supply of this valuable resource. This thesis proposes an SPVP forecasting method that applies DL methodologies using real data from multiple solar power stations. SPVP time series data is complex and characterized by variable, dynamic, and multidimensional attributes. Consequently the research in this thesis has to address various challenges, predominantly stemming from the inherent characteristics of SPVP data. The multifaceted nature of these challenges includes data variability and non-stationarity, where the influence of diverse environmental conditions, seasonal variations, and geographical factors introduces significant fluctuation and unpredictability into the data. To address this variability, forecasting models that have the capability to adapt and predict based on changing patterns are needed. Additionally, the multi-dimensional nature of the inputs required for precise forecasting poses another hurdle. Accurate SPVP generation forecasting models need to integrate multiple types of data, not only historical generation data but also exogenous vari- ables such as weather conditions. Compounding these challenges is the issue of data availability. Many solar installations, especially new ones or those in less-studied regions, do not have the extensive historical data crucial for train- ing robust forecasting models. Traditional machine learning methods often prove inadequate, as they are limited by their dependence on extensive data manipulation and feature engineering, so the requirements for deep domain expertise—capabilities are not always available. These methods struggle to capture and utilize the dynamic interplay between the factors affecting SPVP generation, and this underscores the need for innovative approaches that can navigate these complexities more effectively. Motivated by the limitations of existing forecasting approaches, this research explores innovative DL techniques capable of handling the complexi- ties of SPVP data. To address the challenges posed by data variability, we introduce an aggregated SPVP model with a Wavelet-based-coefficient (Wco- eff) approach that is used for univariate data decomposition to denoise the data. The Wcoeff model redefines the wavelet transform (WT) application to streamline feature extraction. This approach provides a scalable and accurate forecasting solution by mitigating computational complexity yet retaining temporal relationships. Exogenous data is then integrated to enhance forecasting accuracy, and the research addresses the multi-dimensional nature of these inputs through the innovations of the Multilevel Data Fusion and Neural Basis Expansion Analysis (MF-NBEA) model. This model represents a pivotal advance in using DL for SPVP forecasting. Indeed, understanding the most important lagged variables influencing the generation is crucial for refining forecasting models. Given the high dimensionality and evolving nature of the data to be used, a dynamic approach to lagged variable selection and modeling is required. The research develops dynamic feature selection that adjusts to changing conditions and highlights the most predictive variables over time. This adaptability ensures models remain accurate and relevant, even as the underlying data patterns shift. Finally, we introduce a novel methodology that integrates learned knowledge from multiple source domains to address the critical challenges in fore- casting accuracy when data is scarce. This innovative transfer learning approach marks a significant departure from traditional single-source forecasting methods. By leveraging the wealth of data available from already established solar power installations, the new methodology enhances the forecasting model’s ability to predict solar power output in new locations or locations with limited historical data. The essence of the novelty is in the strategic fusion of knowledge from across multiple domains, utilizing advanced techniques such as average weights fusion and evolutionary optimization based fusion. This thesis makes a significant contribution to the field of DL models and renewable energy forecasting by providing scalable, efficient, and adaptable models. The findings underscore the potential for advanced DL techniques to navigate the complexities of SPVP time series data and offer insights that will facilitate the broader integration of solar energy into the power grid. This work opens avenues for future research to enhance model interpretability, explore cross-domain applications of transfer learning, and further optimize models for real-time forecasting applications.49 0Item Restricted A Deep Learning Framework for Blockage Mitigation in mmWave Wireless(Portland State University, 2024-05-28) Almutairi, Ahmed; Aryafar, EhsanMillimeter-Wave (mmWave) communication is a key technology to enable next generation wireless systems. However, mmWave systems are highly susceptible to blockages, which can lead to a substantial decrease in signal strength at the receiver. Identifying blockages and mitigating them is thus a key challenge to achieve next generation wireless technology goals, such as enhanced mobile broadband (eMBB) and Ultra-Reliable and Low-Latency Communication (URLLC). This thesis proposes several deep learning (DL) frameworks for mmWave wireless blockage detection, mitigation, and duration prediction. First, we propose a DL framework to address the problem of identifying whether the mmWave wireless channel between two devices (e.g., a base station and a client device) is Lineof- Sight (LoS) or non-Line-of-Sight (nLoS). Specifically, we show that existing beamforming training messages that are exchanged periodically between mmWave wireless devices can also be used in a DL model to solve the channel classification problem with no additional overhead. We extend this DL framework by developing a transfer learning model (t-LNCC) that is trained on simulated data and can successfully solve the channel classification problem on any commercial-off-the-shelf (COTS) mmWave device with/without any real-world labeled data. The second part of the thesis leverages our channel classification mechanism from the first part and introduces new DL frameworks to mitigate the negative impacts of blockages. Previous research on blockage mitigation has introduced several model and protocol based blockage mitigation solutions that focus on one technique at a time, such as handoff to a different base station or beam adaptation to the same base station. We go beyond those techniques by proposing DL frameworks that address the overarching problem: what blockage mitigation method should be employed? and what is the optimal sub-selection within that method? To do so, we developed two Gated Recurrent Unit (GRU) models that are trained using periodically exchanged messages in mmWave systems. Specifically, we first developed a GRU model that tackled the blockage mitigation problem in single-antenna clients wireless environment. Then, we proposed another GRU model to expand our investigation to cover more complex scenarios where both base stations and clients are equipped with multiple antennas and collaboratively mitigate blockages. Those two models are trained on datasets that are gathered using a commercially available mmWave simulator. Both models achieve outstanding results in selecting the optimal blockage mitigation method with an accuracy higher than 93% and 91% for single-antenna and multiple-antenna clients, respectively. We also show that the proposed methods significantly increases the amount of transferred data compared to several other blockage mitigation policies.17 0Item Restricted Machine Learning Classififiers for Chronic Obstructive Pulmonary Disease Assessment Using Lung CT Data.(Western University, 2024-04-12) Alsurayhi, Halimah; Abbas, SamaniChronic Obstructive Pulmonary Disease (COPD) is a condition characterized by persistent inflammation and airflow blockages in the lungs, contributing to a significant number of deaths globally each year. To guide tailored treatment strategies and mitigate future risks, the Global Initiative for Chronic Obstructive Lung Disease (GOLD) employs a multifaceted assessment system of COPD severity, considering patient's lung function, symptoms, and exacerbation history. COPD staging systems, such as the high-resolution eight-stage COPD system and the GOLD 2023 three staging systems, have been later developed based on these factors. Lung Computed Tomography (CT) is becoming increasingly crucial in investigating COPD as it can detect various COPD phenotypes, such as emphysema, bronchial wall thickening, and gas trapping. Deep learning techniques show promise in leveraging CT imaging to assess the severity of COPD. This thesis uses lung CT data in conjunction with machine learning techniques to classify COPD patients according to these staging systems. For the eight-stage system, both Neural Network and Convolutional Neural Network (CNN) approaches were employed for classification. To develop the Neural Network model, features were extracted from lung CT scans at inspiration and expiration breathing phases, including lung air features and COPD phenotypes features. The CNN model utilized a single lung CT scan at the expiration phase. The GOLD 2023 three staging system involves training separate CNN models using lung CT scans at expiration to predict symptom levels and COPD exacerbation risk. In this thesis, in addition to models trained from scratch, Transfer Learning was also employed to develop models for the eight-stage COPD classification, Symptom level prediction, and exacerbation risk prediction. The developed classifiers demonstrate reasonably high classification performance, indicating their potential for deployment in clinical settings to enhance COPD assessment using image data.13 0Item Restricted ENSEMBLE MACHINE LEARNING IN SPACE WEATHER ANALYTICS(New Jersey Institute of Technology, 2024) Alobaid, Khalid; Wang, JasonThis dissertation addresses several important space weather problems using ensemble learning techniques. An ensemble method combining multiple machine learning models is often more accurate than the individual machine learning models that form the ensemble method. There are several techniques for constructing an ensemble. With in-depth case studies, the dissertation demonstrates the usefulness and effectiveness of ensemble machine learning for space weather analytics, especially for predicting extreme space weather events such as coronal mass ejections (CMEs). The dissertation begins with an ensemble method for predicting the arrival time of CMEs from the Sun to Earth. The proposed method, named CMETNet, combines classical machine learning algorithms such as support vector regression, random forests, XGBoost and Gaussian process regression, along with a deep convolutional neural network (CNN), to perform multimodal learning. The classical machine learning algorithms are used to learn latent patterns from CME features and background solar wind parameters while the deep CNN is used to learn patterns hidden in CME images where the learned patterns are jointly used to make predictions. Experimental results show that CMETNet outperforms existing models, both machine learning based and physics based. Finally, the dissertation presents a fusion method, named DeepCME, to estimate two important properties of CMEs, namely, CME mass and kinetic energy. The DeepCME method is a fusion of three deep-learning models, namely ResNet, InceptionNet, and InceptionResNet. The fusion model extracts features from Large Angle and Spectrometric Coronagraph (LASCO) C2 images, effectively combining the learning capabilities of the three component models to jointly estimate the mass and kinetic energy of CMEs. To the best of current knowledge, this is the first time that deep learning has been used for CME mass and kinetic energy estimations. DeepCME can help scientists better understand CME dynamics. In conclusion, the dissertation showcases many applications of learning techniques including ensemble learning, deep learning, transfer learning and multimodal learning in space weather analytics. The tools and methods developed from the dissertation will make contributions to the understanding and forecasting of CME dynamics and CME geoeffectiveness.16 0Item Restricted Optimising IDS configurations for IoT Networks Using AI approaches(Saudi Digital Library, 2023) Alshahrani, Abdulmonem; John A. ClarkThe number of internet-connected smart objects, known as the Internet of Things (IoT), has increased significantly in recent years. The low cost of manufacturing has enabled a proliferation of smart devices across many tasks and domains. Such devices, however, are typically resource constrained. This has led to the emergence of Low-Power and Lossy Networks (LLNs) which require efficient communication protocols. The Routing Protocol for Low-Power and Lossy Networks (RPL) has been designed for such a purpose. The RPL is the de-facto standard routing protocol for the IoT. Nevertheless, RPL-enabled networks are susceptible to many attacks as these devices are unattended, resource-constrained, and connected via unreliable networks. Deploying Intrusion Detection Systems (IDSs) in such a large and resource-constrained environment is a challenging task. The resource-constrained nature of many devices and nodes restricts what tasks those nodes can realistically expect to perform. There may be a great many choices as to what detection functionality is allocated and where. There are cost/benefit trade-offs between them and inappropriately favouring one over the another may cause an ineffective IDS deployment. In this research, we investigate the use of a metaheuristic- based optimisation method, namely a Genetic Algorithm (GA), to discover optimal IDS placements and configurations for the Low Power and Lossy Networks (LLNs). To the best of our knowledge, this is the first attempt to optimise IDS configurations for emerging and constrained networks while incorporating a wider set of aspects than currently considered. Our approach seeks to optimise and balance detection performance (either detection rate or F1 score), coverage (nodes are monitored by an appropriate number of probes), feasibility cost (nodes host detection functionality within their capability), and deployment cost (seeking to reduce the number of probes deployed). We propose a framework that makes trades-offs between these functional and non-functional constraints. A genetic algorithm-based optimisation approach is developed to address the IDS optimisation task. However, the fitness function is evaluated in part via a computationally expensive simulation. We show how a neural network can be used as a surrogate fitness function evaluation, providing better results more cheaply. Experimental results show that the proposed function approximation is more computationally efficient. Our approximation-based GA system is 1.6 times faster than the corresponding simulation-based GA system. It also gives better results. Furthermore, when used repeatedly to generate candidate placements and configurations the resource costs per generation reduce drastically. The surrogate model is valuable as it significantly reduces the evaluation time and computation. However, generality is still a limitation. Therefore, we propose a transfer-learning Deep Neural Networks (DNNs) approach, that harnesses the experience of previously trained neural networks, to develop a general proxy model for evaluating IDS configurations of variant newly-presented networks more accurately.24 0Item Restricted Transforming Medical Image Segmentation with Enhanced U-Net Architectures and Adaptive Transfer Learning(Saudi Digital Library, 2023-05-25) Albishri, Ahmed; Yugyung, LeeMedical imaging has revolutionized healthcare by enabling accurate diagnosis, treatment planning, and monitoring of various diseases. Various modalities, such as computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound, visualize diverse anatomical structures and pathological conditions. However, challenges arise in medical image segmentation due to increasing complexity, variability, noise, artifacts, and scarcity of annotated data. The advent of AI, particularly deep learning with Convolutional Neural Networks (CNNs), has facilitated significant advancements in medical image segmentation. U-Net, a prominent CNN architecture, provides accurate segmentation results with relatively low training samples due to its encoder-decoder structure with skip connections. In addition, transfer learning further mitigates limitations imposed by scarce labeled data. In this thesis, we develop innovative custom U-Net models with advanced building blocks and transfer learning strategies, such as AM-UNet for human brain claustrum segmentation from MRI scans, TLU-Net for organ and tumor segmentation from CT scans, and OCU-Net for oral cancer tissue segmentation from whole slide images (WSI) stained with Hematoxylin and Eosin (H&E). Furthermore, we introduce the "U-Framework", a comprehensive guide in designing and optimizing U-Net models. This framework encompasses key decisions related to architecture, transfer learning, module selection and fine-tuning, and evaluation strategies. Finally, by comparing our models with state-of-the-art approaches on benchmark datasets, we demonstrate their significant potential to contribute to medical image segmentation.30 0