SACM - United Kingdom
Permanent URI for this collectionhttps://drepo.sdl.edu.sa/handle/20.500.14154/9667
Browse
28 results
Search Results
Item Restricted Deep Learning based Cancer Classification and Segmentation in Medical Images(Saudi Digital Library, 2025) Alharbi, Afaf; Zhang, QianniCancer has significantly threatened human life and health for many years. In the clinic, medical images analysis is the golden stand for evaluating the prediction of patient prog- nosis and treatment outcome. Generally, manually labelling tumour regions in hundreds of medical images is time- consuming and expensive for pathologists, radiologists and CT scans experts. Recently, the advancements in hardware and computer vision have allowed deep-learning-based methods to become main stream to segment tumours automatically, significantly reducing the workload of healthcare professionals. However, there still remain many challenging tasks towards medical images such as auto- mated cancer categorisation, tumour area segmentation, and relying on large-scale labeled images. Therefore, this research studies theses challenges tasks in medical images proposing novel deep-learning paradigms that can support healthcare professionals in cancer diagnosis and treatment plans. Chapter 3 proposes automated tissue classification framework called Multiple Instance Learning (MIL) in whole slide histology images. To overcome the limitations of weak super- vision in tissue classification, we incorporate the attention mechanism into the MIL frame- work. This integration allows us to effectively address the challenges associated with the inadequate labeling of training data and improve the accuracy and reliability of the tissue classification process. Chapter 4 proposes a novel approach for histopathology image classification with MIL model that combines an adaptive attention mechanism into an end-to-end deep CNN as well as transfer learning pre-trained models (Trans-AMIL). Well-known Transfer Learning architectures of VGGNet [14], DenseNet [15] and ResNet[16] are leverage in our framework implementation. Experiment and deep analysis have been conducted on public histopathol- ogy breast cancer dataset. The results show that our Trans-AMIL proposed approach with VGG pre- trained model demonstrates excellent improvement over the state-of-the-art. Chapter 5 proposes a self-supervised learning for Magnetic resonance imaging (MRI) tu- mour segmentation. A self-supervised cancer segmentation framework is proposed to re- duce label dependency. An innovative Barlow-Twins technique scheme combined with swin transformer is developed to perform this self supervised method in MRI brain medical im- ages. Additionally, data augmentation are applied to improve the discriminability of tumour features. Experimental results show that the proposed method achieves better tumour seg- mentation performance than other popular self- supervised methods. Chapter 6 proposes an innovative Barlow Twins self supervised technique combined with Regularised variational auto-encoder for MRI tumour images as well as CT scans images segmentation task. A self-supervised cancer segmentation framework is proposed to reduce label dependency. An innovative Barlow-Twins technique scheme is developed to represent tumour features based on unlabeled images. Additionally, data augmentation are applied to improve the discriminability of tumour features. Experimental results show that the pro- posed method achieves better tumour segmentation performance than other existing state of the art methods. The thesis presents four approaches for classifying and segmenting cancer images from his- tology images, MRI images and CT scans images: unsupervised, and weakly supervised methods. This research effectively classifies histopathology images tumour regions based on histopathological annotations and well-designed modules. The research additionally comprehensively segments MRI and CT images. Our studies comprehensively demonstrate label-effective automatic on various types of medical image classification and segmentation. Experimental results prove that our works achieve state-of-the-art performances on both classification and segmentation tasks on real world datasets8 0Item Restricted Stress Detection: Leveraging IoMT Data and Machine Learning for Enhanced Well-being(Saudi Digital Library, 2025) Alsharef, Moudy Sharaf; Alshareef, Moudywe focus on the detection of acute stress, characterized by short-term physiological changes such as changes in heart rate variability (HRV), breathing patterns, and other bodily functions. Often measurable through wearable or contactless sensors. Accurate detection of acute stress is crucial in high-pressure environments, such as clinical settings, to reduce cognitive overload, prevent burnout, and minimize errors. Current research on stress detection faces multiple challenges. First, most proposed methods are not designed to identify stress in unseen subjects, limiting their generalizability and practical applicability. Second, due to the sensitive nature of stress-related physiological data and the risk of data leakage, insufficient attention has been paid to ensuring data privacy while preserving utility. Third, many existing studies rely on synthetically induced stress in controlled environments, overlooking real-world scenarios where stress can have severe consequences. Finally, nearly all research in this domain employs invasive IoMT sensors or wearable devices, which may not be practical or scalable for real-world applications. This thesis presents five key contributions in the field of stress detection using Internet of Medical Things (IoMT) sensors and machine learning. First, it introduces a deep learning model based on self-attention (Transformer), trained and evaluated using the WESAD dataset, a widely used benchmark collected from 15 participants under controlled stress tasks. The model achieved 96% accuracy in detecting stress and was validated using leave-one-subject-out (LOSO) cross-validation to demonstrate generalizability to unseen individuals. Second, to ensure data privacy, a differential privacy framework was integrated into the model. This approach adds noise during training to prevent sensitive data leakage and achieved 93% accuracy, confirming it is both private and effective. Third, the thesis introduces a new dataset called PARFAIT, collected from 30 healthcare workers during real hospital duties (ICU, ER, OR) using non-invasive HRV sensors and the Maslach Burnout Inventory (MBI) to label stress levels. This dataset supports real-world analysis of stress among physicians. Fourth, a cost-sensitive model is developed using XGBoost and the PARFAIT dataset, assigning higher penalties to stress misclassifications that could lead to medical errors. This model achieved 98% accuracy and reduced false negatives, making it suitable for clinical settings. Finally, a contactless radar-based system is presented to detect stress using ultrawideband (UWB) radar, capturing HRV and breathing data. A deep learning model achieved 92.35% accuracy, offering a non-wearable, scalable alternative. Although the radar-based model achieved a slightly lower accuracy (92.35%) compared to the wearable-based model (96%), it provides several important advantages. It works with out any physical contact, helps maintain user privacy, and can be more practical to deploy in clinical settings where wearable sensors may not be suitable. The small drop in accuracy is mainly due to the limitations of radar in measuring HRV precisely. However, by combining radar-based HRV with breathing features, the overall performance remains competitive. 314 0Item Restricted Improving Induction Motor Fault Classification Accuracy Through Enhanced Multimodal Preprocessing, Artificial Image Synthesis, Deep Learning and Load-Adaptive Graph-Based Methods(Saudi Digital Library, 2024-11-06) Hejazi, Shahd Ziad M; Packianather,, Michael Liu, YingThis thesis aims to improve the accuracy of fault classification in Induction Motor (IM) bearings by developing and applying advanced Artificial Intelligence (AI) and Machine Learning (ML) techniques for condition monitoring data. The proposed framework utilises several approaches, namely, Multimodal Data Preprocessing, Artificial Thermal Image Creation, Customised Radial Load Assessment, Multimodal Systems Decision Fusion, and Graph Convolutional Networks (GCN) on Tabular Datasets to achieve better classification accuracies over existing methods. This study's first significant contribution is the proposed novel approach in the preprocessing of multimodal condition monitoring data for classifying induction motor faults that employs Convolutional Neural Networks (CNNs), such as Residual Network-18 (ResNet-18) and SqueezeNet, to fuse vibration signals and thermal images. This approach enhances fault classification accuracy by 14.81% and proves exceptionally effective in scenarios with compromised image quality. Further refinement using Gramian Angular Field (GAF) processing enhances the detection of subtle fault indicators, achieving better accuracy than Continuous Wavelet Transform (CWT). Secondly, this thesis explores the creation of high-quality artificial thermal images using Wasserstein GAN with Gradient Penalty (WGAN-GP) and its conditional variant, conditional Wasserstein GAN with Gradient Penalty (cWGAN-GP), to address the scarcity of thermal imaging data. The artificial thermal images replicate complex thermal patterns of IMs under various fault conditions with remarkable accuracy, as evidenced by the improved Maximum Mean Discrepancy (MMD) scores and a 40.00% reduction in training times. The high fidelity of these artificially generated images, validated against real images, underscores their practical use in fault classification. Thirdly, the Customised Load Adaptive Framework (CLAF) introduces a novel approach to incorporating load variations into fault classification. Through a two-phase process involving ANOVA and optimal CWT, load-dependent fault subclasses—Mild, Moderate, Severe, and Normal (fault-free) or Healthy—are identified. The CLAF achieved an accuracy of 96.30% ± 0.50% in 18.155 s during five-fold cross-validation using a Wide Neural Network (WNN), demonstrating its ability to detect subtle fault variations across different Load Factors (LFs). Fourthly, building upon the CLAF’s load-dependent fault subclass structure, the research proposed two key methodologies for enhancing load-specific condition monitoring accuracy while optimising training time relative to complexity using the MFPT bearing dataset namely, the Load-Dependent Multimodal Vibration Signal Enhancement and Fusion (LD-MVSEF) method, and the Hybrid Graph-CNN Decision Fusion (HG-CDF) method. The LD-MVSEF employs a multimodal approach across multiple channels, with different signal encoding techniques achieving a fault classification accuracy of 99.04% ± 0.22% over five runs in 18 min 30 s. It performed particularly well in the Moderate class, achieving 99.15% ± 0.89% testing accuracy, and scored 97.20% ± 1.75% in the Mild class. The proposed HG-CDF combines the structural strengths of Graph Convolutional Networks (GCNs) with the pattern-detection capabilities of 1D-Convolutional Neural Networks (1D-CNNs) for CLAF load-dependent fault subclass classification. The study began by optimising the GCN through Taguchi experiments, converting tabular data into graph structures using the k-Nearest Neighbours method and achieving a mean accuracy of 89.01% ± 1.25 across nine configurations. HG-CDF further improved performance, reaching an overall accuracy of 99.19% in just 3 minutes and 28 seconds, surpassing LD-MVSEF in the Mild class with 98.92% accuracy while also providing a faster and more efficient solution. The methodologies proposed in this research significantly enhance the IM fault classification task, improve the decision-making process, and offer scalable solutions adaptable to other domains.10 0Item Restricted Human Action Recognition Based on Convolutional Neural Networks and Vision Transformers(University of Southampton, 2025-05) Alomar, Khaled Abdulaziz; Xiaohao, CaiThis thesis explores the impact of deep learning on human action recognition (HAR), addressing challenges in feature extraction and model optimization through three interconnected studies. The second chapter surveys data augmentation techniques in classification and segmentation, emphasizing their role in improving HAR by mitigating dataset limitations and class imbalance. The third chapter introduces TransNet, a transfer learning-based model, and its enhanced version, TransNet+, which utilizes autoencoders for improved feature extraction, demonstrating superior performance over existing models. The fourth chapter reviews CNNs, RNNs, and Vision Transformers, proposing a novel CNN-ViT hybrid model and comparing its effectiveness against state-of-the-art HAR methods, while also discussing future research directions.31 0Item Restricted Rasm: Arabic Handwritten Character Recognition: A Data Quality Approach(University of Essex, 2024) Alghamdi, Tawfeeq; Doctor, FaiyazThe problem of AHCR is a challenging one due to the complexities of the Arabic script, and the variability in handwriting (especially for children). In this context, we present ‘Rasm’, a data quality approach that can significantly improve the result of AHCR problem, through a combination of preprocessing, augmentation, and filtering techniques. We use the Hijja dataset, which consists of samples from children from age 7 to age 12, and by applying advanced preprocessing steps and label-specific targeted augmentation, we achieve a significant improvement of a CNN performance from 85% to 96%. The key contribution of this work is to shed light on the importance of data quality for handwriting recognition. Despite the recent advances in deep learning, our result reveals the critical role of data quality in this task. The data-centric approach proposed in this work can be useful for other recognition tasks, and other languages in the future. We believe that this work has an important implication on improving AHCR systems for an educational context, where the variability in handwriting is high. Future work can extend the proposed techniques to other scripts and recognition tasks, to further improve the optical character recognition field.56 0Item Restricted Enhance Deep Learning for Cybersecurity Challenges in Software-Defined Networks(University of Warwick, 2024-09) Alsaadi, Sami; Leeson, Mark and Lakshminarayana, SubhashTraditional network devices, such as a router or switch, incorporate the control plane and the data plane. IT operators independently set traffic policies on each device. Nonetheless, this architectural setup raises operational expenses and complicates the dynamic adaptation and maintenance of secure network configurations. Software-defined Networking (SDN) represents a revolutionary approach to network management, offering enhanced flexibility. SDN promotes rapid innovation in networking by centralizing control and making it programmable. However, security concerns pose significant barriers to the broader adoption of SDN, as this new architecture potentially opens novel attack vectors previously non-existent or more challenging to exploit. Machine Learning (ML) strategies for SDN security rely heavily on feature engineering, requiring expert knowledge and causing delays. Therefore, enhancing intrusion detection is essential for protecting SDN architectures against diverse threats. The thesis develops techniques for detecting malicious activities in SDN using Deep Learning DL. It starts by evaluating CNNs on an SDN dataset, leading to a new CNN-based detection approach that employs a novel regularization method to reduce kernel weights and address overfitting, improving effectiveness against unrecognized attacks. Additionally, a semi-supervised learning method using an LSTM autoencoder combined with One Class SVM is introduced, specifically designed to detect DDoS attacks. This approach enhances the detection capabilities within SDN environments, showcasing the potential of DL in advancing network security.26 0Item Restricted Enhancing Breast Cancer Diagnosis with ResNet50 Models: A Comparative Study of Dropout Regularization and Early Stopping Techniques(University of Exeter, 2024-09-20) Basager, Raghed Tariq Ahmed; Kelson, Mark; Rowland, SarehEarly detection and treatment of breast cancer depend on accurate image analysis. Deep learning models, particularly Convolutional Neural Networks (CNNs), have proven highly effective in automating this critical diagnostic process. While prior studies have explored CNN architectures [1, 2], there is a growing need to understand the role of dropout regularization and fine-tuning strategies in optimizing these models. This research seeks to improve breast cancer diagnosis by evaluating ResNet50 models trained from scratch and fine-tuned, with and without dropout regularization, using both original and augmented datasets. Assumptions and Limitations: This research assumes that the Kaggle Histopathologic Cancer Detection dataset is representative of real-world clinical images. Limitations include dataset diversity and computational resources, which may affect generalization to broader clinical applications. ResNet50 models were trained on the Kaggle Histopathologic Cancer Detection dataset with various configurations of dropout, early stopping, and data augmentation [3–6]. Performance was assessed using accuracy, precision, recall, F1-score, and AUC-ROC metrics [7, 8]. The best-performing model was a ResNet50 trained from scratch without dropout regularization, achieving a validation accuracy of 97.19%, precision of 96.20%, recall of 96.90%, F1-score of 96.55%, and an AUC-ROC of 0.97. Grad-CAM visualizations offered insights into the model’s decision-making process, enhancing interpretability crucial for clinical use [9,10]. Misclassification analysis showed that data augmentation notably improved classification accuracy, particularly by correcting previously misclassified images [11]. These findings highlight that training ResNet50 without dropout, combined with data augmentation, significantly enhances diagnostic accuracy from histopathological images. Original Contributions: This research offers novel insights by demonstrating that a ResNet50 model without dropout regularization, trained from scratch and with advanced data augmentation techniques, can achieve high diagnostic accuracy and interpretability, paving the way for more reliable AI-powered diagnostics.20 0Item Open Access Parking Occupancy Classification: Deep learning model compression for edge device classification(Queen Mary University of London, 2024) Tamim, Ziad; Ansari, Tayyab AhmedUrban areas face severe traffic congestion due to poorly managed parking systems. Advanced parking management, like automated and smart parking guidance systems, offers a feasible solution requiring real-tim occupancy data. Traditional sensor-based methods are costly and inefficient for large scale parking, whereas video-based sensing is more effective. Deep learning methods improve accuracy but have high computational costs, affecting real-time performance. Central servers or cloud platforms are often used but can be impractical due to high resource demands. Instead, utilising edge devices with model compression techniques—such as quantisation and knowledge distillation enhances efficiency. This study aims to boost the inference speed of parking classification algorithms by developing a custom model called QCustom based on the MobileNet Depthwise Separable Convolution blocks and using compression techniques to reduce the inference time further. Contributions include developing an edge-based real-time occupancy system, setting performance benchmarks, optimising models for edge devices, and testing on a prototype parking lot. The goal is efficient and accurate parking management for smart cities. Results of the paper include accuracy of 98.8% on the CNRPark-EXT dataset, real world implementation accuracy of 97.44%, and an inference speed for one parking slot of 0.746ms on the Raspberry Pi 5.46 0Item Restricted LIGHTREFINENET-SFMLEARNER: SEMI-SUPERVISED VISUAL DEPTH, EGO-MOTION AND SEMANTIC MAPPING(Newcastle University, 2024) Alshadadi, Abdullah Turki; Holder, ChrisThe advancement of autonomous vehicles has garnered significant attention, particularly in the development of complex software stacks that enable navigation, decision-making, and planning. Among these, the Perception [1] component is critical, allowing vehicles to understand their surroundings and maintain localisation. Simultaneous Localisation and Mapping (SLAM) plays a key role by enabling vehicles to map unknown environments while tracking their positions. Historically, SLAM has relied on heuristic techniques, but with the advent of the "Perception Age," [2] research has shifted towards more robust, high-level environmental awareness driven by advancements in computer vision and deep learning. In this context, MLRefineNet [3] has demonstrated superior robustness and faster convergence in supervised learning tasks. However, despite its improvements, MLRefineNet struggled to fully converge within 200 epochs when integrated into SfmLearner. Nevertheless, clear improvements were observed with each epoch, indicating its potential for enhancing performance. SfmLearner [4] is a state-of-the-art deep learning model for visual odometry, known for its competitive depth and pose estimation. However, it lacks high-level understanding of the environment, which is essential for comprehensive perception in autonomous systems. This paper addresses this limitation by introducing a multi-modal shared encoder-decoder architecture that integrates both semantic segmentation and depth estimation. The inclusion of high-level environmental understanding not only enhances scene interpretation—such as identifying roads, vehicles, and pedestrians—but also improves the depth estimation of SfmLearner. This multi-task learning approach strengthens the model’s overall robustness, marking a significant step forward in the development of autonomous vehicle perception systems.43 0Item Restricted A Peer-to-Peer Federated Learning Framework for Intrusion Detection in Autonomous Vehicles(Lancaster University, 2024-09) Alotaibi, Bassam; Bradbury, MatthewAs autonomous vehicles (AVs) increasingly rely on interconnected systems for enhanced functionality, they also face heightened cyberattack vulnerability. This study introduces a decentralized peer-to-peer federated learning framework to improve intrusion detection in AV environments while preserving data privacy. A novel soft-reordering one-dimensional Convolutional Neural Network (SR-1CNN) is proposed as the detection engine, capable of identifying known and unknown threats with high accuracy. The framework allows vehicles to communicate directly in a mesh topology, sharing model parameters asynchronously, thus eliminating dependency on centralized servers and mitigating single points of failure. The SR-1CNN model was tested on two datasets: NSL-KDD and Car Hacking, under both independent and non-independent data distribution scenarios. The results demonstrate the model’s robustness, achieving detection accuracies of 94.39% on the NSL-KDD dataset and 99.97% on the Car Hacking dataset in independent settings while maintaining strong performance in non-independent configurations. These findings underline the framework’s potential to enhance cybersecurity in AV networks by addressing data heterogeneity and preserving user privacy. This research contributes to the field of AV security by offering a scalable, privacy-conscious intrusion detection solution. Future work will focus on optimizing the SR-1CNN architecture, exploring vertical federated learning approaches, and validating the framework in real-world autonomous vehicle environments to ensure its practical applicability and scalability.55 0
- «
- 1 (current)
- 2
- 3
- »