Electrical and Computer Engineering

Permanent URI for this collectionhttps://hdl.handle.net/10315/35603

Browse

Now showing 1 - 20 of 102

Access status: Open Access ,
Design and Validation of an Instrumented Knee Brace with Adjustable Restrictions for Machine Learning-Based Gait Analysis
(2026-07-24) Dhupalia, Rahim Hussainali; Grau, Gerd
Gait data can be analysed to classify which activities are performed and gauge the gait-related health of the performer. We consider health in the context of knee osteoarthritis and gait-afflicting conditions. This work captures gait from sixteen healthy participants who performed walking and sit-to-stand activities in a lab while artificially restricted at four severities, using a custom collection device and instrumented brace. A convolutional neural network classified activities with participant-averaged accuracies of 65.61% ± 8.85% and 67.21% ± 9.06% on two well-performing hyper parameter sets. Analysis found time-series features that differentiated between some restriction levels for each of the activities. Significant difference was found between swing phase of unrestricted walking activities and all other restriction levels, and for toe-off of unrestricted walking activities and moderate and severe restrictions. For sit-to-stand, significant difference was found between jerk standard deviation of negotiating sitting edge of unrestricted data and moderate and severe restrictions.
Access status: Open Access ,
Large Language Models (LLMs) for Wireless Network Resource Management and Forecasting Applications
(2026-07-24) Khan, Muhammad Umar; Tabassum, Hina
Next-generation wireless networks require intelligent methods that can both optimize resource allocation and anticipate future network behaviour under rapidly changing conditions. Conventional optimization and deep learning approaches have demonstrated strong performance in specific wireless tasks, but they often rely on expensive retraining, large task-specific datasets, and fixed model assumptions that limit adaptability across new environments. This thesis investigates whether large language models (LLMs) can serve as training-free, prompt-engineered reasoning agents for wireless systems, with the goal of developing a unified methodology that adapts to diverse wireless problems without gradient-based retraining. The thesis first focuses on radio resource management (RRM), where constrained non-convex optimization problems arise in tasks such as beamforming, power control, and user association. To address these problems, this thesis proposes LLM4RRM (LLM for Radio Resource Management), a multi-agent LLM framework inspired by block coordinate descent, in which specialized solver agents optimize subsets of decision variables while an evaluator agent iteratively refines their outputs through structured feedback. An LLM-guided adaptive differentiable projection mechanism is introduced to improve feasibility under coupled constraints, together with a prompt-based uncertainty quantification framework for assessing solution reliability and consistency. The thesis then extends the same training-free agentic design philosophy to multivariate time-series forecasting through LLM4TSF (LLM for Time-Series Forecasting), where specialized agents model temporal components of correlated signals and an evaluator agent enforces statistical consistency across multiple series. Overall, the results demonstrate that prompt-engineered multi-agent LLM systems can provide adaptable, feasible, and scalable wireless intelligence without task-specific retraining.
Access status: Open Access ,
Energy Efficient Arrhythmia Classifier Based on Event-Driven Data
(2026-07-24) Zhao, Yixiao; Lian, Yong Peter
Electrocardiogram (ECG) monitoring is important for early detection of cardiac arrhythmias, but traditional uniformly sampled ECG acquisition often leads to high power consumption in long-term wearable systems. Level-crossing (LC) sampling provides an event-driven representation that can reduce redundant data. This thesis investigates heartbeat classification using LC ECG signals. First, a feature-based classification framework is developed. ECG fiducial points are detected from LC signals, and morphological, rhythm, and LC dynamic features are constructed. After feature optimization, a multilayer perceptron classifier achieves an accuracy of 99.4% with a Macro-F1 of 0.949 on the MIT-BIH database. Second, an event-driven spiking neural network (SNN) classifier is proposed. A time-aware leaky integrate-and-fire neuron model is introduced to incorporate event intervals into membrane dynamics. The proposed SNN achieves 99.02% accuracy and 0.936 Macro-F1, while improving the recognition of fusion beats. The results demonstrate the potential of event-driven ECG classification for low-power wearable monitoring systems.
Access status: Open Access ,
Optimal Deployment of Energy Storage Technologies for Grid Service Provision and Microgrid Formulation
(2026-07-24) Ahmed, Yahia Nabil Abdellatif; Farag, Hany E. Z.
The global transition towards deep decarbonization and the widespread integration of renewable energy resources introduce unprecedented operational complexities within modern power systems. Managing these dynamics necessitates massive flexibility across both transmission and distribution networks, which can be achieved through Energy Storage (ES) and decentralized microgrid architectures. Despite the rapid advancement of these technologies, significant barriers remain in their effective integration into the power grid, including the complex modeling of emerging technologies, stringent market regulations, and the lack of electrically informed decentralization strategies. This thesis presents comprehensive and novel optimization frameworks to address these critical gaps in the deployment of ES technologies for Grid Services (GSs) provision and the optimal formulation of sustainable microgrids. At the transmission level, a holistic techno-economic framework is proposed for the optimal sizing and multi-horizon scheduling of grid-connected Advanced Adiabatic Compressed Air Energy Storage (AA-CAES). By incorporating detailed thermodynamic constraints and diverse electricity pricing schemes, the model assesses AA-CAES participation across Energy Arbitrage (EA), Operating Reserve (OR), and Capacity Markets Demand Response (CMDR). Results indicate that while AA-CAES is economically viable for EA and OR primarily in high-variability markets, it remains highly profitable for CMDR across both stable and variable pricing markets. At the consumer level, the research develops an optimal energy resource sizing and management framework for Large Energy Facilities (LEFs) participating in CMDR. To overcome the severe computational complexities introduced by dynamic baseline evaluation rules, a novel convex relaxation of CMDR constraints is introduced. This approach reduces computation time from several hours to fractions of a second while ensuring global optimality with an optimality gap below 0.1\%. The framework further demonstrates that aggregating short-duration (e.g., Battery Energy Storage (BES)) and long-duration (e.g., Hydrogen Energy Storage (HES) and Compressed Air Energy Storage (CAES)) technologies maximizes cost savings and environmental sustainability through enhanced arbitrage, emissions-cutting peak-demand reduction, and sustained capacity market obligations. At the distribution network level, a tri-stage optimization framework is designed to systematically decentralize Power Distribution Systems (PDSs) into self-sustaining microgrids. This framework introduces a hierarchical agglomerative clustering method based on electrical line impedance, followed by the stochastic planning of Distributed Energy Resources (DERs). A subsequent allocation strategy optimally places active and reactive power resources, significantly enhancing voltage regulation and reducing system losses under both islanded and cooperative operational modes. These proposed frameworks provide robust, computationally tractable solutions to maximize the economic and operational benefits of ES and decentralized microgrids in modern power systems.
Access status: Open Access ,
Beyond Functional Correctness: Training-Time and Inference-Time Approaches to Improving Non-Functional Quality of LLM-Generated Code
(2026-07-24) Sivapiran, Sanjeepan; Uddin, Gias
Large Language Models (LLMs) generate functionally correct code that often lacks software standard qualities such as security, readability, and maintainability. This thesis investigates complementary approaches to bridge this gap. The first study examines training-time alignment using DPO and BoNBoN across five LLMs and their instruction-tuned variants. Results show non-functional alignment achieves consistent improvements (10.6\% average) while functional alignment proves unreliable (4.9\% average), with effectiveness varying by model family and pathway. The second study introduces POSec, an inference-time framework combining automatic prompt optimization with Selective Prompt Anchoring for secure code generation. Evaluated across 8 LLMs, 5 languages, and 3 optimizers, POSec achieves +20–33\% security improvements at Pass@10 without model retraining. Together, these studies demonstrate that improving LLM-generated code quality requires multi-level interventions, with training-time alignment offering breadth and inference-time optimization offering targeted depth.
Access status: Open Access ,
Optimizing Log Parsing: Template Refinement and Semantic Generalization
(2026-07-24) Enan, Nafid Ajmain; Uddin, Gias
Logs are a primary source of runtime information in modern software systems and are fun- damental to tasks such as anomaly detection, failure diagnosis, performance monitoring, and reliability engineering. Log parsing, the process of converting unstructured log messages into structured templates, serves as the foundation of these downstream analyses. To assist in this, many novel approaches have been proposed. This dissertation contains three sections: 1) an empirical study of existing log parsers, 2) SynLog+, and 3) GeLT. In the empirical study, we evaluate syntax-based, semantic-based, and hybrid log parses. The observations from the study indicates two issues. Firstly, there is a gap between the pars- ing accuracy of the syntax- and semantic-based parsers that hybrid parsers aim to bridge. However, semantic-based and hybrid parsers leverage PLMs or LLMs, requiring heavy run- time costs. Secondly, semantic-based parsers lack in generalization ability over unseen logs. With SynLog+, we propose a wrapper to existing log parsers that acts as a template refinement module aiming to improve the parsing accuracy without added runtime cost. With GeLT, we aim to re-evaluate the existing generalizable log parser, Log3T, and identify its design flaws and propose an optimized generalizable semantic-based log parser.
Access status: Open Access ,
A Scalable Hardware-Efficient CSNN Accelerator with Binarized STDP and STDG-based Hybrid Learning
(2026-07-24) Rahimian Kalatehbali, Hamid; Amirsoleimani, Amirali
Spiking neural networks (SNNs) offer a biologically inspired, energy-efficient computing paradigm that processes information through sparse, event-driven spike signals. In this work, we present Binarized spike-timing dependent plasticity-gradient (BSTDPG), a novel SNN framework that combines Time-to-first-spike (TTFS) latency encoding with simple integrate-and-fire (IF) neuron model, binarized convolutional layers with local Spike-timing dependent plasticity (STDP) learning rule, and a gradient-based classifier. To support hybrid learning, the final output layer is trained using either a Support Vector Machine (SVM) or gradient-based backpropagation, allowing flexible deployment in both unsupervised and supervised settings. The proposed architecture consists of two spiking convolutional layers followed by maxpooling and a classifier. A key innovation is the zero-skip optimization technique, which dynamically prunes inactive spike regions based on bounding boxes generated from TTFS-encoded spike trains. This drastically reduces unnecessary computations during training and inference. For instance, processing sample digit '0' during its first timestep without optimization results in full 28 × 28 convolution and 7860 ns latency. With zero-skip enabled, the latency drops to 1710 ns, achieving a 4.6× speedup and 78.2% processing time reduction. A similar improvement is observed at the full network level, where zero-skip reduces prediction latency from 1459 µs to 1212 µs, saving 246 µs. We evaluate BSTDPG on Modified National Institute of Standards and Technology database (MNIST) and Fashion-MNIST (F-MNIST) across both shallow and deep network configurations. The model achieves accuracies of 99.1% and 91.0%, respectively, outperforming or matching several state-of-the-art SNNs such as BANN, BS4NN, S4NN, and even deep temporal SNNs, despite using fewer layers. Together, these innovations highlight BSTDPG as a promising architecture for low-latency, low-power neuromorphic applications including edge-AI, robotics, and event-based vision, offering an effective trade-off between biological plausibility, efficiency, and learning performance.
Access status: Open Access ,
Laser-Induced Graphene Electrochemical Sensors: Plasma Treatment, Lithium-Ion Detection via Printed Surface Modification, and Integrated 3D Printed Microfluidics
(2026-07-24) Shovan, Saumik Dey; Grau, Gerd
Laser-induced graphene (LIG) has gained attention as a cost-effective and sustainable alternative to traditional electrochemical sensing electrodes. It enables a simple, one-step fabrication process that eliminates the need for complex materials or procedures. While LIG has attracted significant attention for various sensing applications, its potential for lithium detection remains largely unexplored. Lithium detections are particularly noteworthy given that monitoring is an essential priority in both industrial processes and medical diagnostics. One known limitation of LIG is its hydrophobic surface, which limits sensor performance and was addressed in the first part of this study. Although plasma treatment has been used to improve LIG’s surface wettability, the effect of treatment duration has not been fully explored. Overexposure can introduce surface defects and reduce conductivity. We optimized plasma treatment time to improve surface reactivity while preserving the structural and electrical properties of LIG. In the second part of this study, we developed a lithium-selective sensor to address the increasing demand for reliable lithium detection. To achieve lithium selectivity, we functionalized the LIG electrode with lithium manganese oxide (LMO) using dispense printing for better uniformity of the printed layer. LMO is chosen for its strong interaction with lithium ions. This thesis is the first report of LMO being combined with LIG for lithium sensing. Finally, this study introduced a fully additively manufactured electrochemical sensor that integrated functionalized LIG electrodes with 3D-printed microfluidic channels. By enabling both the formation of conductive LIG patterns and the 3D printing of microfluidic structures directly onto a 3D-printed substrate, this approach removes the need for complex alignment or assembly steps. The successful result is a compact, efficient, and scalable sensing platform that combines the strengths of LIG, lithium-selective functionalization, and additive manufacturing for practical and selective lithium detection.
Access status: Open Access ,
Mixed Graph Signal Analysis of Joint Image Denoising and Interpolation
(2026-07-24) Viswarupan, Niruhan; Cheung, Gene
An image signal captured by a photonic sensor is corrupted by noise and other distortions, and requires several operations before rendering a presentable photo, including denoising, demosaicking, rectification, and white balancing. Specifically, denoising reduces acquisition noise without destroying image details, while interpolation generates novel pixels at new 2D locations using acquired image data. Such image operations are typically applied in sequence---e.g., denoising an image first followed by interpolation---which is sub-optimal in general. In this thesis, we jointly optimize denoising and interpolation operations from a graph signal processing (GSP) perspective. Specifically, we first develop two theorems that show one-to-one mappings between (pseudo-)linear denoisers / interpolators and undirected / directed graph filters that are solutions to variational optimization problems regularized using different graph smoothness priors. Leveraging these theorems, we investigate scenarios where joint denoising / interpolation operations would outperform separate operations in sequence, and mathematically derive those optimal joint operators. Experiments demonstrate validation results, where optimized joint operators outperformed separate operators in sequence in various practical imaging scenarios.
Access status: Open Access ,
Design and Optimization of Sensor Interface Circuitry and All-Digital PLL in 22nm Technology with gm/ID Methodology Using Precalculated Look-Up Tables.
(2026-03-10) Mohammadpanah, Amirhossein; Magierowski, Sebastian
This thesis presents the design of a monolithic System-on-Chip (SoC) front-end for solid-state nanopore DNA sequencing, implemented in GlobalFoundries 22nm FD-SOI technology. A major bottleneck in portable sequencing is the massive parasitic capacitance of the sensor (approx. 5 pF), which severely limits the bandwidth required to detect rapid, picoampere-level ion currents. Additionally, designing high-performance analog interfaces in deep-submicron nodes using traditional, iterative SPICE simulations is highly inefficient. To address these challenges, this research establishes a systematic, hierarchical "Inverse Design" framework based on the gm/ID methodology. By utilizing 4D Look-Up Tables (LUTs) extracted from BSIM-IMG compact models, this flow deterministically synthesizes transistor geometries directly from system-level specifications. Crucially, the methodology integrates back-gate biasing (FBB) as a fourth design dimension to dynamically trade off leakage for speed and maximize voltage headroom. Utilizing this synthesis flow, a Transimpedance Amplifier (TIA) was designed, achieving a 1 MHz bandwidth and a 5 Mohms transimpedance gain while driving the large capacitive sensor load. To provide the low-jitter timing infrastructure required for full system integration and precise ADC sampling, a wide-range All-Digital Phase-Locked Loop (ADPLL) featuring a MASH 1-1-1 Sigma-Delta Modulator was also implemented, achieving a phase noise of -84 dBc/Hz at a 1 MHz offset. Ultimately, this work demonstrates that a deterministic LUT-based synthesis approach successfully enables robust, high-performance mixed-signal IC design for biosensors while eliminating the reliance on trial-and-error manual sizing.
Access status: Open Access ,
An IR-UWB Transmitter IC with Integrated Lossless Neural Data Compression for High-Channel-Count Neural Implants
(2026-03-10) Korosteliov, Maxim; Kassiri, Hossein
This thesis presents the design, development, and validation of a novel impulse radio ultra-wideband (IR-UWB) transmitter architecture with integrated lossless data compression, targeting high density neural implants. The proposed design incorporates (a) a greatest common divisor (GCD) based compression algorithm that is tailored to the sparse output of a level crossing ADCs (LCADC), (b) an edge combination based UWB transmitter circuit capable of pulse generation for 2-PPM modulation, and (c) programmability to mitigate process, voltage, and temperature (PVT) variations while ensuring spectral compliance. Our results demonstrate the system’s capability to perform real time lossless compression, achieving 50% to ~86% reduction in transmitted bits and up to 94% reduction, leading to significant reductions in transmission energy and enabling scalable channel counts under strict implant power budgets. When combined with the inherent data-rate reduction of LC-ADCs, the overall compression reaches ~259× (corresponding to 99.6% reduction compared to a traditional 10-bit ADC with a 20 kHz sampling rate), enabling proportionally higher channel counts within the same power and bandwidth budget. The compression algorithm was implemented in System Verilog and synthesized in TSMC 180 nm CMOS, achieving 44.7 µW power consumption and 4114 µm2 area for the compression block (from synthesis results). The transmitter was realized using programmable current-starved inverter delays, glitch generators, and a unit-cell amplifier, co-designed with a circular patch antenna centered at 4 GHz. Both the digital controller and the transmitter were fabricated using the TSMC 180nm CMOS process. The digital controller and transmitter were tested with a custom receiver architecture.
Access status: Open Access ,
Anomaly Detection and Attack Mitigation in Federated Learning
(2026-03-10) Huang, Haoqi; Wang, Ping
Federated Learning (FL) enables collaborative model training across multiple participants without sharing raw data, achieving a balance between privacy protection and data utility. However, the decentralized nature of FL also introduces new security threats, among which model poisoning attacks are the most critical. Malicious clients can upload manipulated model updates to disrupt the global aggregation process, leading to performance degradation or even system failure. This thesis systematically investigates anomaly detection (AD) and attack mitigation in FL, aiming to enhance system robustness and security from both detection and defense perspectives. First, a comprehensive review and categorization of AD methods are presented, focusing on reconstruction-based and prediction-based deep learning frameworks and their applications to different data types. Second, existing defense mechanisms are analyzed in terms of their effectiveness and limitations under various attack scenarios, providing the theoretical foundation for the proposed framework. Based on these analyses, this thesis proposes an unsupervised defense framework named Dual-VAE with Truncated Gaussian (DVTG). The framework follows a three-stage structure to model and filter client updates. In Stage 1, a variational autoencoder (VAE) is trained to estimate reconstruction errors and identify a set of potentially clean updates. In Stage 2, a second VAE with a truncated Gaussian prior is trained on this refined subset to obtain a more stable latent representation. In Stage 3, the trained model evaluates incoming client updates and filters those with high reconstruction errors before aggregation. The method enables effective anomaly detection without requiring labeled or clean data and remains stable under both adversarial and stochastic disturbances. Experiments conducted on the MNIST dataset under non-independent and identically distributed (non-IID) conditions show that DVTG outperforms the baseline model across different attack scenarios. The framework effectively detects malicious clients while maintaining stable convergence and comparable accuracy to the non-attack scenario. Finally, this thesis discusses several future directions. These include extending the defense framework to hierarchical FL architectures and developing more interpretable and efficient AD models. The goal is to build a reliable and practical FL framework with stronger defense capability and better adaptability to real-world environments.
Access status: Open Access ,
Single-Stage Interleaved Photovoltaic (PV) Step-up Resonant Converter Modules with Integrated PV MPPT for Magnetron Application
(2025-11-11) Molaei, Aref; Lam, John
With the growing need for compact, efficient, high-voltage DC power supplies for magnetron applications and the global interest in renewable energy sources, the power electronics industry has new challenges and opportunities. Photovoltaic (PV) energy, as a clean energy source, is widely available and requires efficient power conversion interfaces to satisfy the magnetron load's high-voltage, low-ripple, and high-efficiency demands. In this thesis, a high-gain single-stage DC/DC converter with soft-switching capability is proposed for PV-driven magnetron applications. This thesis is divided into two main sections. The first part addresses the topology development and analysis of the proposed converter. The topology combines an integrated interleaved boost stage, a CLL resonant network and a voltage doubler rectifier for high voltage gain and low input current ripple. The converter maintains zero-voltage switching (ZVS) for all the switches and zero-current switching (ZCS) for all the diodes over a wide range of operation conditions. In-depth theoretical analysis and design considerations followed by simulation and experimental verification with a hardware prototype, are presented in this section. The second part of the thesis offers maximum power point tracking (MPPT) control schemes tailored for PV integration. Two duty-ratio-based approaches are utilized: a discrete logic-based Perturb and Observe (P&O) algorithm and a closed-loop proportional-integral (PI) controller. A small-signal modeling of the boost-integrated stage is addressed to ensure dynamic stability and accurate tracking under variable irradiance conditions. The converter’s performance and the effectiveness of the control strategies are validated through both simulation and experimental testing.
Access status: Open Access ,
Convergent Active Stereo: Investigating Depth Perception with a Robotic Binocular Camera System
(2025-11-11) Chi, Mingshi; Tsotsos, John K.
While extensive research exists on stereo disparity estimation algorithms for fixed parallel systems, there is limited exploration of active convergent stereo cameras. This thesis investigates the differences in depth perception between a robot head mimicking human eye movement with the ability to fixate and a static parallel setup. A novel convergent stereo algorithm, which is capable of searching for correspondences with diagonal epipolar lines, is proposed. To evaluate the algorithm, the first natural image convergent stereo dataset was created, which includes ground truth for parallel views, parallel images of each scene, and images of fixations across the scene. An in-depth analysis reveals the advantages and limitations of both systems, highlighting the geometric benefits of active convergent stereo, especially in challenging scenes. The results from these findings can contribute to further the development of more effective vision-based depth perception system designs for humanoid robots.
Access status: Open Access ,
Interpretable Deep Image Denoiser by Unrolling Graph Laplacian Regularizer
(2025-11-11) Hosseini, Seyed Alireza; Cheung, Gene
Image denoising is a fundamental problem in image restoration. Researchers have studied the problem for decades and proposed numerous algorithms. In the past ten years, deep learning has produced complex models that deliver high-quality denoised images, but these models require on large numbers of parameters, lack interpretability, and depend heavily on random parameter initialization. As a result, they frequently converge to poor-performing local minima. This thesis proposes an image denoising neural net constructed by unrolling an iterative algorithm solving a maximum a posteriori (MAP) optimization problem regularized using a graph Laplacian prior. To guarantee a minimum level of performance, we initialize the network to a known (pseudo-)linear denoiser, which is mapped to a corresponding graph Laplacian matrix specifying the MAP problem, leveraging a previous linear algebraic theorem. The performance of the network is further enhanced by learning an appropriate perturbation matrix to augment the graph Laplacian via a lightweight convolutional neural net (CNN). This design bridges the gap between classical model-based methods with modern deep learning, eliminates the need for random initialization, reduces parameter count, and improves interpretability of the constructed network. Experiments show that our method demonstrates competitive image quality reconstruction compared to state-of-the-art deep learning models, while offering improved robustness, interpretability, and parameter efficiency.
Access status: Open Access ,
Hardware Accelerated Basecalling for Mobile Nanopore DNA Sequencing
(2025-11-11) Beyene, Abel; Magierowski, Sebastian; Ghafar-Zadeh, Ebrahim
Nanopore Sequencing is characterized by its high-throughput and long reads which makes it amenable to sequencing genomes in real-time. The vast amounts of data generated, however, imposes serious requirements on any hardware tasked with performing the back-end data processing necessary to convert the electrical signal inputs to base-pair label outputs. The literature contains several examples of machine learning algorithms targeted at nanopore sequencing and demonstrated on server class CPUs and GPUs. Such hardware would be poorly suited to mobile environments however which have tight power and area constraints. To demonstrate the potential of custom accelerators for sequencing tasks, this work presents a complete RISC-V System-on-Chip (SoC) equipped with a bioinformatics hardware accelerator and taped out in Global Foundries 22nm process. It’s able to achieve a 13X speed-up when basecalling over the ARM cortex A53 while only consuming 20 mW power.
Access status: Open Access ,
Deep Learning-Based Detection, Annotation and Staging of Breast Cancer Metastasis on Whole-Slide Histopathology Images of Lymph Nodes
(2025-07-23) Tauqeer, Abdullah; Sadeghi-Naini, Ali
Breast cancer is one of the most frequently diagnosed malignancies and a leading cause of cancer-related mortality among women worldwide. Accurate detection and characterization of breast cancer lymph node metastasis are crucial for effective patient management. In this thesis, two novel deep learning frameworks have been developed and investigated to automate the analysis of hematoxylin and eosin (H&E) stained digital pathology whole-slide images (WSIs) of lymph nodes to streamline the breast cancer pathology workflow in the clinic. First, a nuclei segmentation and classification model, TexSegNet, is introduced to extract detailed cellular-level information from the digital pathology images of biopsied lymph nodes. This model has been trained and evaluated using 7,904 histology images from the PanNuke dataset, which includes 205,343 annotated nuclei across 19 tissue types (2,351 images from breast tissue). TexSegNet leverages a hybrid encoder-decoder architecture, integrating multi-scale convolutions, nuclear texture extraction blocks, advanced attention mechanisms, and a feedback-driven classification branch. Trained on all tissue types of the PanNuke dataset and subsequently fine-tuned on its breast subset, TexSegNet has achieved an overall accuracy of 81.4±0.4% in detecting and classifying nuclei on the breast test set, substantially outperforming established benchmarks such as HoVer-Net, CellViT, and StarDist. Notably, TexSegNet has maintained consistently good performance across various cell types, including underrepresented ones, with F1-scores of 89.3±0.4%, 91.1±0.5%, 88.9±0.8%, and 84.3±0.3% in detecting and classifying neoplastic, epithelial, inflammatory, and connective cell nuclei, respectively. Second, a selective neighborhood attention-based multiple instance learning framework (MIL) is proposed for automated detection, localization, and staging of lymph node metastases in breast cancer. The proposed framework leverages a dual-path feature extractor, incorporating both nuclei segmentation/classification outputs and transformer-based tissue features, alongside a dynamic attention mechanism that selects and emphasizes neighboring patches based on similarity to the target patch. Trained and optimized on 269 WSIs (over 14 million image patches) from the CAMELYON16 dataset and evaluated on an independent test set of 129 WSIs, the model achieves 96.2±1.5% sensitivity, 95.3±2.4% precision, and 95.7±3.1% F1-score in patch-level classification, along with the area under the receiver operating characteristic curve (AUC) of 0.96±0.01 for slide-level classification. External validation on an out-of-distribution (OOD) subset of 30 annotated WSIs from the CAMELYON17 dataset (representing multiple institutions) demonstrates robust generalizability, with a patch-level F1-score of 87.0±1.8% and a slide-level AUC of 0.88±0.03. Furthermore, for 500 WSIs from 100 patients in the CAMELYON17 dataset, the framework demonstrates an excellent performance in patient-level pN-staging, with a quadratic-weighted Cohen’s kappa of 0.94±0.02, indicating near-expert concordance in detecting and classifying the extent of nodal metastasis. Ablation analyses underscore the importance of incorporating nuclei-based features and selective neighborhood attention, with noticeable performance degradation observed when either element is removed. By integrating insights into histologic type, morphology and spatial heterogeneity at the cellular level with contextual information at the tissue level, the proposed framework effectively replicates key aspects of human pathological assessment and shows promise as a decision-support tool for metastatic breast cancer detection, annotation, and staging in the era of digital pathology. Together, these two frameworks, TexSegNet for detailed nuclei segmentation and classification, and the selective neighborhood attention-based MIL model for targeting metastatic tumor detection and staging, form an integrated pipeline for breast cancer pathology assessment on standard H&E-stained slides. Together, they demonstrate strong potential to reduce pathologist workload, reduce subjectivity in diagnosis, and serve as processing tools in digital pathology to derive quantitative biomarkers for further analysis or prognostication. Future work will focus on expanding dataset diversity, enhancing model explainability, and integrating these algorithms into clinical workflows to support precision oncology care.
Access status: Open Access ,
Transfer Learning for Data-Driven Power Flow and Optimal Power Flow Applications
(2025-07-23) Nazari, Dorsa; Srikantha, Pirathayini
Ensuring real-time grid operations is essential for maintaining both stability and efficiency in today’s dynamic power systems. While machine learning (ML)-based approaches enable fast inference, these models are often trained using datasets derived from static grid configurations, such as fixed topologies. Adapting these models to evolving grid conditions introduces additional complexity and necessitates acquiring supplementary datasets, which require computationally intensive solvers. This paper presents a method that improves computational efficiency compared to conventional transfer learning techniques for adapting ML models used in power flow (PF) and optimal power flow (OPF) analysis to changing grid conditions. Our findings indicate that only 6% of the original dataset requires recalibration, and the entire process of data point regeneration and model fine-tuning is completed in under 3 seconds in the benchmark IEEE 14-bus, IEEE 118-bus, and PEGASE 1352-bus systems.
Access status: Open Access ,
Field Testing and Economic Analysis of Residential Vehicle to Grid Deployment
(2025-07-23) Nasr, Khunsha; Farag, Hany E. Z.
The development of Electric Vehicles (EVs) has surged in recent years, positioning them as direct replacements for fossil fuel-dependent internal combustion engines. However, the rise in EVs will place considerable strain on the grid, necessitating transmission and distribution infrastructure upgrades. In this regard, EVs equipped with bidirectional charging can act as independent energy storage, managing energy at home during normal and emergency conditions, feeding excess energy back into the grid to reduce its strain, and potentially generating revenue for their owners. However, the high cost of Residential Bidirectional Chargers (RBC) and the limited availability of EVs that support bidirectional charging, particularly Vehicle-to-Home/Grid (V2H/G), remain significant barriers. Despite extensive theoretical research on bidirectional charging of EVs, there is a lack of real-world testing and comprehensive techno-economic analysis to assess the feasibility of the wide deployment of these technologies in residential areas. This thesis aims to investigate the techno-economic viability for the wide deployment of RBC throughout the following: 1. Field Testing for RBCs: Conducting practical evaluations of RBCs to gather empirical data on their performance, efficiency, and reliability under real-world conditions. 2. Identification of Policy and Regulatory Barriers: Analyzing the legislative and regulatory frameworks in Ontario to identify obstacles that hinder the deployment of RBCs, and proposing solutions to overcome these barriers. 3. Development of a Mathematical Model: Creating a model to perform a regional-wide cost-benefit analysis for RBC deployment programs, considering three main stakeholders: local distribution companies, EV owners, and ratepayers. By addressing these aspects, this research will provide a comprehensive understanding of the practical and economic implications of implementing RBC technologies, paving the way for its broader adoption and integration into the energy infrastructure.
Access status: Open Access ,
ARC-C: Analytical Framework and Software Tool for Automated Risk-Based Cryptoperiod Calculation in Industrial Control Systems
(2025-04-10) Cianfarani, Gabriele Alberto; Vlajic, Natalija
Over the past decade, industrial control systems (ICSs) and critical infrastructure (CI) have become prime targets for advanced persistent threat (APT) groups and nation-state actors due to their potential for severe impact. This has resulted in the cybersecurity community increasing their focus on ICS/CI threat modelling and defence. This thesis examines the crucial role of the internal network reconnaissance stage of ICS/CI attacks, particularly those using the OPC UA standard with encrypted in-transit data. We first introduce a comprehensive attack tree outlining data siphoning strategies and highlight the importance of periodic encryption-key rotation to mitigate risk. Noting the lack of clear cryptoperiod guidelines in industry standards, we then present the Automatic Risk-based Cryptoperiod Calculation (ARC-C) framework. ARC-C aims to optimally determine cryptoperiod lengths based on security risks and operational constraints. We demonstrate its application in two realistic ICS environments: a Water Treatment Plant and an Energy Storage System.

Browse

Recent Submissions