Electrical and Computer Engineering

Permanent URI for this collectionhttps://uwspace.uwaterloo.ca/handle/10012/9908

This is the collection for the University of Waterloo's Department of Electrical and Computer Engineering.

Research outputs are organized by type (eg. Master Thesis, Article, Conference Paper).

Waterloo faculty, students, and staff can contact us or visit the UWSpace guide to learn more about depositing their research.

Browse

Now showing 1 - 20 of 2082

Strain-Balanced InGaAs/InAlAs Superlattices on InP(111)B for Terahertz Photoconductive Antennas
(University of Waterloo, 2025-05-27) Hosseini Farahabadi, Seyed Ali
Terahertz time-domain spectroscopy (THz-TDS) systems hold great promise for next generation communication, imaging, sensing, and metrology applications, all of which demand stringent performance requirements. Extending the deployment of THz-TDS systems beyond research laboratories into practical field applications requires the development of cost-effective and portable systems compatible with 1550 nm femtosecond fiber-coupled lasers. InGaAs/InAlAs photoconductors grown on InP(001) have emerged as strong candidates for efficient THz signal generation and detection. However, further material modifications are necessary to optimize their performance. One promising strategy involves growing InGaAs/InAlAs superlattices (SLs) along crystal orientations other than [001], such as [111], which can significantly influence their electronic and optical properties. Despite this potential, the growth and characterization of InGaAs/InAlAs SLs on (111)-oriented InP substrates remain underexplored due to intrinsic growth complexities and financial constraints. Addressing this gap, the present work investigates the molecular beam epitaxy (MBE) growth of InGaAs/InAlAs SLs on InP(111)B substrates. Through rigorous theoretical design, controlled MBE growth, and detailed structural and electrical characterization, we successfully achieved an atomically smooth, strain-balanced InGaAs/InAlAs SL on InP(111)B at 450°C for the first time. Leveraging the polar nature of the [111] orientation and strain engineering, simulations revealed a strong piezoelectric field of 153 kV/cm across the SL. This systematic approach enabled a detailed analysis of how structural parameters— such as indium composition and layer thickness—affect carrier dynamics, evaluated using time-resolved pump-probe spectroscopy at 1550 nm. Complementary absorption measurements indicated an enhanced absorption coefficient reaching 5195 cm⁻¹, while Hall effect characterization showed carrier mobility as high as 2756 cm²/Vs. These findings mark a crucial step toward achieving low-temperature-grown InGaAs/InAlAs structures with subpicosecond carrier lifetimes. While high-quality SLs were realized at 450°C, the impact of lower growth temperatures on structural quality remained unclear and warranted further investigation. To this end, two 50-period InGaAs/InAlAs SLs—one nominally lattice-matched and the other strain-balanced—were grown on InP(111)B substrates, with the growth temperature systematically reduced from 450°C to 200°C. Cross-sectional scanning transmission electron microscopy (STEM) revealed that lowering the temperature from 450°C to 400°C led to the formation of various defects and stacking faults within the SL grown at 400°C. Further temperature reduction resulted in spatial modulation of interfaces, the formation of microtwins, and phase separation in both the InGaAs and InAlAs layers. A comparative study of strain-balanced SLs grown on (001)- and (111)-oriented substrates under identical conditions showed that while high crystalline quality could be preserved on InP(001), maintaining structural integrity on InP(111) requires careful temperature-specific optimization. Building on these findings, we evaluated the sensitivity of InGaAs and InAlAs crystal quality to the temperature using two novel SLs with modulated growth temperature profiles. In the first structure, InGaAs and InAlAs layers were grown at 450°C and 200°C, respectively; in the second, the respective growth temperatures were reversed. The first SL exhibited a well-defined structure, although the InGaAs-on-InAlAs interfaces appeared slightly more diffuse than the InAlAs-on-InGaAs counterparts. These results suggest that growing InGaAs at 450°C can help mitigate interface roughness introduced by low-temperature InAlAs growth, thereby preserving SL integrity throughout the growth process. Despite the structural imperfections observed in SLs grown entirely at low temperatures on InP(111)B, these structures remained functional for THz photoconductive applications. This promising behavior led to a further investigation of carrier dynamics using strain-balanced InGaAs/InAlAs SLs grown on stationary substrates with varied indium compositions in the wells and barriers. Remarkably, trapping times as short as 1 ps and carrier lifetimes as fast as 4 ps were achieved without relying on complex Be doping schemes. Collectively, these advancements provide new insight into the controlled growth of InGaAs/InAlAs SLs on InP(111)B substrates, paving the way for the development of high performance electronic and photonic devices operating at telecommunication wavelengths.
Strain-Balanced InGaAs/InAlAs Superlattices for 1550 nm-based Terahertz Photoconductive Antennas
(University of Waterloo, 2025-05-26) Entezami, Milad
Terahertz (THz) technology has developed as a transformative tool with critical applications in spectroscopy, imaging, and high-speed communication across diverse fields. Recent advancements have facilitated a transition from conventional photoconductors, driven by laboratory-scale and costly Ti:Sapphire lasers operating at 800 nm wavelengths, toward compact, cost-effective, and industrially viable THz time-domain spectroscopy systems utilizing a fiber-optic platform. Despite these advances, developing high-performance photoconductive materials remains a significant challenge, especially for efficient THz generation and broadband detection at the telecom wavelength of 1550 nm. This research investigates the growth, characterization, and photoconductive properties of strain-balanced InGaAs/InAlAs superlattices grown on InP substrates as promising photoconductive materials for fiber-compatible THz photoconductive sources and detectors. First, fundamental principles underlying photoconductivity, including carrier generation, transport, and recombination, are reviewed, highlighting their critical role in determining overall performance. Recent material engineering strategies targeting THz operation at telecom wavelengths are also discussed. Despite recent advancements, there remains a need for photoconductive materials simultaneously exhibiting enhanced optical absorption, superior carrier transport, and ultrafast recombination lifetimes to enable efficient broadband THz operation. Based on these principles, a theoretical framework describing carrier dynamics for transient THz photocurrent generation and detection is reviewed. This theoretical foundation is complemented by comprehensive experimental studies, including time-resolved pump-probe spectroscopy, Hall effect measurements, optical absorption spectroscopy, and band structure simulations, enabling precise quantification of carrier lifetime, mobility, and optical absorption coefficients. Molecular beam epitaxy was employed to implement a stationary (non-rotating) substrate growth method, enabling precise lateral control of structural parameters across the wafer. The primary contribution lies in strategically balancing compressive strain within InGaAs wells with tensile strain in InAlAs barriers, thereby achieving a net-zero stress condition in each superlattice period. This strain-balancing approach systematically addresses limitations inherent to conventional lattice-matched epitaxy, resulting in significantly enhanced crystal quality in strain-induced superlattices, improved optical absorption at 1550 nm, and optimized electronic transport properties. Experimental results confirm these structural optimizations improve carrier dynamics, essential for high-performance THz devices. Moreover, low-temperature-grown Be-doped strain-balanced superlattices are comprehensively characterized, revealing sub-picosecond carrier recombination lifetimes, increased mobility, and enhanced optical absorption. Detailed structural analyses through high-resolution X-ray diffraction, atomic force microscopy, and transmission electron microscopy identify notable Be-induced interdiffusion effects at interfaces. Lastly, modulation doping strategies are explored to further refine photoconductive properties. By systematically controlling dopant distribution within the superlattice layers, this study reveals the complex interplay among doping, defect formation, and strain, enabling precise tuning of carrier transport and recombination dynamics critical for advanced THz photoconductive applications. This research advances the fundamental understanding of carrier dynamics and transport in strain-balanced photoconductive superlattices and offers practical guidance for developing high-performance, telecom-compatible THz photoconductive materials well-suited for portable pulsed THz spectroscopy and imaging systems.
Electrochemical Capacitance-Voltage Profiling of Carrier Distributions in Advanced III-V Semiconductor Epitaxial Structures
(University of Waterloo, 2025-05-20) Qudsi, Yazan
This thesis presents a comprehensive investigation of electrochemical capacitance-voltage (ECV) profiling for the characterization of carrier concentration profiles in III-V semiconductor heterostructures, with a focus on GaAs, InAlAs, and GaN. The methodology of ECV is demonstrated in detail, including electrolyte preparation, surface etching mechanics, and data interpretation. ECV profiling of staircase structures is used to calibrate doping concentrations during epitaxial growth, enabling precise evaluation of growth parameters. In low-temperature grown (LTG) GaAs, undoped layers exhibit n-type behavior attributed to excess As antisites, consistent with prior deep-level defect studies. Parabolic Quantum Well (PQW) and Si-δ-doped structures are analyzed, with carrier profiles compared directly against nextnano simulations to address total available carriers and doping accuracy. Electrolyte comparisons show that 0.2M NaOH/0.1M EDTA/10% vol. ED yields superior etch uniformity, while 0.1M Tiron offers sharper resolution with trade-offs in etched surface quality. The study confirms ECV as a valuable diagnostic and calibration tool for advanced semiconductor device development and doping control.
High-Frame-Rate Ultrasound Characterization of Carotid Pulse Waves to Assess Cerebrovascular Resistance
(University of Waterloo, 2025-05-16) Hsu, Yi Han
Objective: Devise an ultrasound imaging framework for cerebrovascular resistance assessment by characterization of carotid pulse waves. Background: The resistance of cerebrovasculature regulates blood flow to the brain that could serve as a biomarker for detection of early dementia. The onset of dementia, leading to higher cerebrovascular resistance (CVR), is theorized as the cerebrovascular damage due to elevated pulse pressure, one heartbeat at a time. The increased pulse pressure causes mechanical stress to the cerebral microvasculature as it propagates into the brain and alters cerebral hemodynamics. This change in cerebral hemodynamics can be explained from the classical Ohm’s law, whereby resistance is the ratio of potential difference (pressure) to current (blood flow). In fact, early dementia patients were identified with higher CVR in several brain regions. CVR can be assessed by measuring the pulse pressure wave and blood flow wave present in the carotid artery during each cardiac contraction. The forward pulse wave propagates along the carotid artery to the brain and is partially reflected back to the heart when encountering a change in resistance in the brain. Higher CVR affects the reflected pulse wave and, accordingly, alters the measured pulse wave in the carotid artery with distinct characteristics such as greater amplitude, broader peak, and higher pulse wave velocity. By examining the forward and reflected pulse wave, pressure and flow information can be acquired and in turn assess CVR. Proposed Solution: To assess CVR from the carotid, an ultrasound imaging framework is proposed due to its low-cost and high accessibility compared to fMRI and PET. This ultrasound framework is developed based on high-frame-rate ultrasound (HiFRUS) paradigm with frame rate up to 10k frame per second. HiFRUS enables the estimation of blood flow velocity and the capture of transient pressure dynamics in the carotid artery where conventional ultrasound cannot achieve. To realize the innovation, four research modules will be pursued: (1) to separate the pulse waves into forward and reflected pulse wave, a wave separation algorithm is developed. (2) to characterize the separated pulse waves, a sensitive analysis was performed by manipulating downstream resistance in vitro study. (3) to transit in vitro to in vivo study, a motion-compensation algorithm was developed. (4) to assess CVR by the pulse wave in the carotid artery, an in-vivo study will be conducted. Impact: This thesis establishes the feasibility of assessing CVR through our proposed pulse wave analysis platform, providing a new method for researchers to investigate new possibilities in the dementia fields. Future work will benchmark our approach against the established imaging methods.
An Empirical Study of Logging Practice in CUDA-based Deep Learning Systems
(University of Waterloo, 2025-05-15) Chen, An
Although logging practices have been extensively explored in conventional software systems, there remains a lack of understanding of how logging is applied in CUDA-based deep learning (DL) systems, despite their growing adoption in practice. In this paper, we conduct an empirical study to examine the characteristics and rationales of logging practices in these systems. We analyze logging statements from 33 CUDA-based open-source DL projects, covering both general-purpose logging libraries and DL-specific logging frameworks. For each type, we identify the development or execution phases in which the logs are used, investigate the reasoning behind their usage, and the relationship between the two different types of logging. Our quantitative analysis reveals that the majority of logging statements occur during the model training phase, with significant usage also in the model loading phase and model evaluation/validation phase. We also observe that logging is predominantly used for monitoring purposes and tracking model-related information. Furthermore, we found that complementary is the most prevalent relationship between general and DL-specific logging. Our findings not only shed light on current logging practices in CUDA-based DL development but also provide practical guidance on when to use DL-specific versus general-purpose logging, helping practitioners make more informed decisions and guiding the evolution of DL-focused logging tools to better support developer needs.
Design of a Multi-Stage Power Amplifier for a 16 Element MIMO Transmitter Testbed
(University of Waterloo, 2025-05-14) Sancak, Bulent Ege
The next-generation standard for fifth-generation (5G) wireless communication demands significant advancements in the transmitter front end compared to its predecessor. This need arises from the exponential increase in both data rate requirements and the number of connected devices. The primary approach to achieving these improvements relies on the adoption of massive multiple-input multiple-output (mMIMO) systems. These systems, combined with beamforming technology, which enhances the equivalent isotropic radiated power (EIRP) by focusing transmission power in specific directions, increase the data rate of the communication systems. While 5G aims to leverage millimeter-wave frequencies for increased bandwidth and data rates, the sub-6 GHz spectrum remains valuable, offering a practical balance between range and performance. Given the new performance requirements, the power amplifier (PA) must exhibit high linearity to minimize memory effects and distortion. As the most power-intensive component in the transmitter chain, the PA also faces increasing efficiency challenges, particularly due to the rising peak-to-average power ratio (PAPR) in 5G systems. In mMIMO configurations, additional complexities arise from factors such as load mismatch and antenna crosstalk, which further impact PA performance. Just as important, electromagnetic interference (EMI) within the transmission chain itself can degrade overall system efficiency —an often overlooked concern, which is the issue that this thesis addresses. To address these challenges, this thesis presents a multi-stage Class AB PA operating in the 3.2–3.8 GHz range, designed for linearization within a 4×4 mMIMO transmitter array. External aluminum shielding is also employed to counteract the EMI. The PA is evaluated through S-parameter, continuous-wave (CW), and modulated signal simulations. CW simulations indicate that the driver and PA together achieve a small-signal gain of approximately 23–25 dB. The PA, utilizing 6W MACOM CGHV1F006 transistors, reaches saturation at 37 dBm output power. The 1 dB compression point is observed around 36 dBm, providing a broad linearity range before saturation. The PA-stage demonstrates power-added efficiency (PAE) between 54% and 61% at maximum power, with a maximum phase distortion of -4 degrees at high-power levels. Modulated signal simulations, conducted with a 100 MHz modulation bandwidth, confirm that the PA is linearizable under single-input single-output (SISO) digital predistortion (DPD). The application of DPD reduces the adjacent channel power ratio (ACPR) from -35 dBc to -55 dBc, demonstrating a significant improvement in linearity compared to the non-DPD case.
Disturbance Vulnerability Analysis and Reduction for Nonlinear Systems using Modes of Instability
(University of Waterloo, 2025-05-08) Wang, Jinghan
Engineered systems naturally experience large disturbances which have the potential to disrupt desired operation because the system may fail to recover to a desired exponentially stable equilibrium point (SEP). It is valuable to determine the mechanism of instability when the system is subject to a particular finite-time disturbance, because this information can be used to improve vulnerability detection, and to design controllers capable of mitigating these disturbances to enhance system resilience and ensure reliable performance, thereby reducing vulnerability. For a large class of nonlinear systems there exists a particular unstable equilibrium point (UEP) on the region of attraction (RoA) boundary of the desired SEP such that the unstable eigenvector of the Jacobian at this UEP represents the mode of instability for the disturbance. Unfortunately, it is challenging to find this mode of instability, especially in high dimensional systems, because it is often computationally intractable to compute this particular UEP for a given disturbance. Consider a particular finite time disturbance applied to a nonlinear system which possesses a SEP representing desired behavior. The system is able to recover from the disturbance if its post disturbance initial condition (IC) lies inside the RoA of the desired SEP. In cases where the system fails to recover, the nonlinear mode of instability for the disturbance represents the subset of system dynamics most responsible for this failure to recover. This thesis develops a novel algorithm for numerically computing the mode of instability for parameter-dependent nonlinear systems without prior knowledge of the particular UEP, resulting in a computationally efficient method. The key idea is to first consider the setting where the system recovers, and to average the Jacobian along the system trajectory from the post-disturbance state up until the Jacobian becomes stable. As the system approaches inability to recover, the averaged Jacobians converge to the Jacobian at the particular UEP, and can be used to extract the unstable eigenvector representing the mode of instability. Convergence guarantees are provided for computing the mode of instability, both for the theoretical setting in continuous time, and for the proposed algorithm which relies on numerical integration. Numerical examples illustrate the successful application of the method to identify the mechanism of instability in power systems subject to temporary short circuits. Then a novel approach to control design for reducing disturbance vulnerability of nonlinear systems using knowledge of the mode of instability is developed. The main idea is to tune controller parameter values so as to drive the post-disturbance IC further inwards away from the RoA boundary by driving it in the direction opposite to the mode of instability. To achieve this, the problem is formulated as a nonconvex optimization problem, and an efficient algorithm is developed to solve it. Local convergence guarantees are provided for this method. Numerical examples illustrate the successful application of the method to reduce the vulnerability of power systems subject to temporary short circuits.
Discriminating and Localizing Thermal Aging in Low Voltage Polymeric Cables using Non-Destructive Electrical Diagnostics
(University of Waterloo, 2025-05-05) Banerjee, Sarajit
Currently, in North America and Europe, significant attention is being paid to the condition of LV electrical cable systems in nuclear power plants (NPPs), given industry and governmental efforts to extend the operational life of numerous existing NPPs well beyond their initial 40-year design life / license period. Thermal degradation is a key cause of long-term aging of LV NPP polymeric cable insulation materials in normally dry areas, and in-situ condition monitoring techniques form an important means of diagnosing its deleterious effects within the scope of a LV cable aging management program. However, key knowledge gaps have thus far precluded the development and deployment of quantitative assessment criteria or predictive modelling/classification methods which can be practically applied towards in-situ, electrically-based diagnostics of thermally aged LV NPP cables across varying insulation types, environmental conditions, rates of dielectric aging, and aging volumes/damage ratios. This thesis investigates fundamental and practical aspects of applying a combined electrical diagnostic approach using DC time-domain, low frequency (0.001 – 1000 Hz) AC, and travelling wave (RF) based in-situ terminal electrical diagnostic techniques to discriminate and localize thermal aging in LV NPP cables. A unique custom experimental set-up is designed and constructed to subject 197 XLPE and EPR prepared LV shielded cable test specimens (prepared from full-scale LV NPP cables) to accelerated thermal aging, whilst simultaneously considering the influence of varying, field-representative experimental conditions. Representative diagnostic data based on approximately 2120 low frequency dielectric spectroscopy (LFDS)/polarization depolarization current (PDC) and 1175 frequency domain reflectometry (FDR) measurements on these test specimens are subjected to varying forms of data analytics including empirical analysis of full spectra, partial least squares regression based predictive modelling, supervised ensemble decision tree-based (random forest) classification and predictor importance analysis. Experimental results from AC/DC (LFDS, PDC-based) electrical diagnostic tests, reference physical-chemical material characterization tests and related data analytics illustrate that material and dielectric changes due to accelerated thermal aging can be discriminated using LFDS and PDC-based electrical diagnostics in combination with interpretable empirical, statistical or supervised ML-based techniques. It is however important to derive a suitably wide range of physically meaningful potential predictor metrics based on the underlying full spectrum dielectric data, including real permittivity (ε′), imaginary permittivity (ε″), polarization current and depolarization current. It is also imperative to undertake all analyses on an insulation-specific basis for LV NPP cables given differences in polymer formulations and physical-chemical evolutions with accelerated thermal aging, and undertake quantitative, insulation-specific approaches to interpret which predictors bear higher responsibilities for the observed dielectric data variance and trends (with respect to thermal aging). Experimental results from RF (FDR-based) electrical diagnostic tests and related data analytics show that thermal aging can be localized using differential, inverse fast Fourier transform (IFFT) converted S11 responses, in combination with advanced window-based signal processing methods and interpretable, supervised ML-based binary anomaly classification techniques. The research supports the premise that FDR measurements should be utilized for defect localization (ideally based on long-term monitoring/trending) rather than condition assessment, especially for samples in early stages of thermal degradation (e.g., antioxidant depletion, start of mechanical hardening) as present in this research. The research outcomes highlight the importance of utilizing an experimental design and broad population dataset for LV NPP cable aging management research which at minimum, considers field-representative variations in thermal aging rates, thermal aging volume/damage ratio, thermal aging locations, measurement (ambient temperatures), and insulation type. Adopting such an approach in this work has allowed for a robust, insulation-specific assessment of critical diagnostic predictors for thermal aging discrimination and localization, and proof-of-concept development and validation of predictive/classification-based models which can yield conservative performance assessments for new (i.e., previously unseen) diagnostic data obtained from field-representative populations. The research undertaken can thus be considered a key knowledge development step to facilitate the deployment of AC, DC and RF electrical non-destructive examination (NDE) techniques towards the aging management of thermally degraded LV cables installed in NPPs or other critical environments.
Advanced AI for Histopathological Whole Slide Image Classification and Captioning
(University of Waterloo, 2025-05-01) Raju , S M Taslim Uddin
Microscopic assessment of histopathology images is vital for accurate cancer diagnosis and treatment. Classification and captioning Whole Slide Images (WSIs) for pathological analysis is an essential but not extensively explored aspect of computer-aided pathological diagnosis. Challenges arise from insufficient datasets and the effectiveness of model training. Generating automatic caption reports for various gastric adenocarcinoma images is another challenge. Moreover, microscopic WSIs face challenges such as redundant patches and unknown patch positions due to subjective pathologist captures. The thesis is divided into two folds. At first, a hybrid method referred to as TransUAAE-CapGen is introduced to generate histopathological captions from WSI patches. The TransUAAE-CapGen architecture consists of a hybrid UNet-based Advereasrial Autoencoder (AAE) for feature extraction and a transformer for caption generation. The hybrid UNet-based AAE extracted complex tissue properties from histopathological patches, transforming them into low-dimensional embeddings. The embeddings are then fed into the transformer to generate concise captions. The second fold focuses on how to reduce redundant patches and handle the unknown patch positions due to subjective pathologist captures. To address these challenges, a novel GNN-ViTCap framework is introduced for classification and caption generation from histopathological microscopic images. A visual feature extractor is used to extract feature embeddings. The redundant patches are removed by using deep embedded clustering to dynamically cluster the images and extracting representative images through a scalar dot attention mechanism. The graph is formed by constructing edges from similarity matrix, ensuring that each node is connected to its closest neighbors. Therefore, graph neural network is utilized to extract and represent contextual information from both local and global areas. The aggregated image embeddings are then projected into the language model’s input space using a linear layer and combined with input caption tokens to fine-tune the large language models for caption generation. Our proposed methods are validated using the BreakHis and PatchGastric microscopic datasets. Several ablation studies have been performed to validate our methods. Experimental analysis demonstrates that the proposed TransUAAE-CapGen and GNN-ViTCap architectures outperform state-of-the-art approaches. Our approaches can be integrated into clinical workflows to facilitate early diagnosis and treatment planning, ultimately enhancing patient care using whole slide images.
AdaptPrompt with Diffusion Set: A Unified Framework for Generalizable Deepfake Detection
(University of Waterloo, 2025-04-30) Jiang, Yichen
Deepfake detection focuses on identifying synthetic media. It has critical applications in cybersecurity, misinformation mitigation, digital forensics, and media authentication. Re- cent developments in deepfake detection have achieved impressive performance, leveraging deep learning models to distinguish between real and synthetic content. However, recent developments in diffusion-based generative models and publicly available tools such as Stable Diffusion and DALL-E pose immediate challenges to existing detection techniques. Diffusion models generate photorealistic and high-resolution content with less observable or detectable artifacts and are, therefore, more difficult to detect using traditional deepfake detection techniques. In this thesis, we present a comprehensive study of existing deepfake detection tech- niques and adapting large vision-language models, i.e., CLIP, to generalizable deepfake detection. We introduce the Diffusion Set, a new dataset of 100k diffusion-generated fake images and 100k real images. Our experiments reveal that detectors trained on Diffusion Set outperform detectors trained on GAN-based datasets. To further enhance deepfake detection, we introduce a new transfer learning strategy that learns randomly initialized prompts and a lightweight adapter network while having CLIP frozen. Extensive experi- ments confirm its efficiency, and we also investigate the impact of dropping specific CLIP layers on detection accuracy. Our study utilizes the Diffusion Set for training and evaluates models on 25 unseen test sets, covering images synthesized by GAN-based models, diffusion-based models, and commercially available tools. Beyond large-scale training, we assess model performance in few-shot settings, where models are trained with only a small fraction of the dataset (e.g., 320 real and 320 fake images), providing insights into their adaptability under data constraints. Additionally, we extend our analysis beyond classification by exploring image attribu- tion and training models in a few-shot setting to attribute images to specific generators such as BigGAN, StarGAN, and Stable Diffusion. Our findings showcase the robustness of CLIP-based models in deepfake detection and their ability to generalize across unseen gen- erative techniques. We also investigated the feasibility of using the same transfer learning strategy for attribution, with experimental results that demonstrate its high effectiveness in closed set attribution.
Investigation, Fabrication, and Characterization of Metal-Insulator-Semiconductor InGaN/GaN Micro-LEDs with Superior Efficiency
(University of Waterloo, 2025-04-29) Yin, Jian
Since the invention of high-efficiency Indium Gallium nitride/Gallium nitride (InGaN/GaN) light-emitting diodes (LEDs), they have become the most widely used visible light sources in the first two decades of the 21st century. Despite the “efficiency droop” problem, InGaN/GaN LEDs offer key advantages for illumination and display applications in electronics, including high brightness, high efficiency, and long lifetime. As the size of mobile electronics decreases, high-resolution displays and miniaturized light sources must also scale down. InGaN/GaN micro-LEDs were considered the most promising candidates for meeting the market demand. However, in recent years, micrometer-scale organic LEDs (OLEDs) have dominated the portable electronics market. The main limitation of InGaN/GaN micro-LEDs is their low efficiency, particularly at low working current densities. Unlike the damage-free inkjet printing techniques used to pattern OLEDs, the dry etching techniques required for patterning InGaN/GaN micro-LEDs cause unavoidable sidewall damage and reduce efficiency. Nonetheless, InGaN/GaN micro-LEDs retain incomparable advantages, such as high-speed response and long lifespan, which provide strong motivation to develop high-efficiency InGaN/GaN micro-LEDs. In this thesis, the mechanisms and solutions to address the low efficiency of InGaN/GaN micro-LEDs are discussed. Chapter 1 provides an overview of the fundamental knowledge of InGaN/GaN heterostructures. Chapter 2 introduces the widely used ABC model to establish the relationship between the Shockley-Read-Hall (SRH) recombination and the internal quantum efficiency (IQE) of micro-LEDs. The properties of surface defects in InGaN/GaN heterostructures are studied. More importantly, nearly all reported practical fabrication techniques, which can effectively suppress surface recombination, are summarized and assessed for applicability. Chapter 3 investigates a widely-neglected mechanism, defects-assisted tunneling recombination, which needs be further studied for InGaN/GaN micro-LEDs. Temperature-dependent external quantum efficiency (EQE) measurements demonstrate that surface defects-assisted tunneling recombination becomes predominant at low temperatures as device size decreases. Quantitative analysis reveals that the role of surface defects-assisted tunneling recombination cannot be overlooked in InGaN/GaN micro-LEDs, even at room temperature. In Chapter 4, a unique device structure is proposed to achieve high-efficiency InGaN/GaN micro-LEDs. A metal-insulator-semiconductor (MIS) structure is fabricated on the sidewall of the device mesa, enabling the application of sidewall bias to adjust the sidewall surface potential of the micro-LED. Quantitative analysis demonstrates that the variation in surface recombination velocity is proportional to the sidewall bias applied to the MIS micro-LEDs. Further calculations indicate that the increase or decrease in IQE is proportional to the sidewall bias and remains constant at low current densities with a fixed sidewall bias. MIS micro-LEDs with 10 μm diameter mesas, fabricated by our collaborator, Vuereal, show promising but modest EQE improvements. The simulated data from numerical simulations match the measured data, verifying the efficiency enhancement of micro-LEDs with MIS structures. The limited efficiency improvement observed in these devices suggests some constraints in the current implementation of MIS micro-LEDs. To fully demonstrate the advantages of the MIS structure, further optimization of both the device structures and fabrication processes is necessary. Therefore, the fabrication and characterization of the MIS micro-LEDs with superior EQE performance are proposed in Chapter 5. Compared to the MIS micro-LEDs provided by our collaborator, the high-efficiency MIS micro-LEDs address the poor ohmic p-contacts problem, replace the insulator in the MIS structure with a 100 nm layer of Al2O3, and employ advanced fabrication methods detailed in Chapter 2. The current density-voltage (JV) characteristics of MIS micro-LEDs with various dimensions show remarkably low leakage current densities, indicating that surface recombination has been minimized. Furthermore, the increase or decrease in EQE of MIS micro-LEDs is observed to be approximately proportional to the magnitude of the positive or negative sidewall bias applied to the MIS micro-LEDs in the range of -20 V to +20 V. This phenomenon is well explained by the analytical model discussed in Chapter 4 and 5. The EQE of the MIS micro-LED with a mesa dimension of 8 μm can be enhanced from 20% to 30.7% (an increase of 10.7%) at a low injection current density of 0.625 A/cm2 by applying a +20 V sidewall bias, which is comparable to the performance of state-of-the-art OLEDs. The maximum EQE of the 8 μm MIS micro-LED was measured to be ~53.9% at an injection current density of 23.3 A/cm2, which is the highest reported EQE to the best of our knowledge.
Breast Cancer Detection Using Microwave Low-Frequency Metasurface-Based Techniques
(University of Waterloo, 2025-04-28) Hernandez Vivanco, Mauricio Enrique
Breast cancer has become an important global public health issue and is the leading cause of death among women. Early detection plays a pivotal role in the effective treatment of breast cancer. Microwave modalities offer a promising alternative because they allow deep penetration into the breast, resulting in an adequate contrast between the electrical properties of the malignant and benign breast tissue. Microwaves are also a mature technique that is relatively affordable and non-ionizing, thus addressing many of the deficiencies of traditional methods. In the first part of this study, we use computer biomechanics finite element simulation software (FEBio) to simulate the physical compression of anatomically realistic breast phantoms. The phantoms are 3D anatomically realistic and made up of MRI images. The compressed breast models derived from this part of this study will create a realistic scenario for the electromagnetic simulation in CST software. We successfully compressed a set of realistic phantoms, and these accurate compressed models created several natural scenarios to test the microwave technology developed in this investigation. The second part of this work introduces an electromagnetic energy scanning technique with a specific application to detect breast cancer. The technique is based on a field detector probe inspired by the concept of a metasurface composed of an ensemble of electrically small elements where a single element represents the field detector. This novel probe provides a resolution that cannot be matched by either electrically small probes or resonance-based probes that have dimensions comparable to the wavelength. The field emanating from a specific structure, such as a human female breast, can be scanned by the proposed probe to achieve a resolution in the millimeter range while operating in the low-microwave frequency spectrum. The probe was tested numerically, and a prototype was tested experimentally, demonstrating its effectiveness in providing a field resolution of approximately 5 mm. The third part of this thesis introduces a novel metasurface sensor designed to detect breast cancer by operating at sub-microwave frequency. Each cell of the metasurface gives rise to a voltage difference that can be measured at the back of the metasurface. Thus, the metasurface has the ability to generate an impression that relates to the constituents of the breast. Simulations were performed to capture and analyze a collection of images using artificial intelligence methodologies. The findings of this testing stage demonstrated the device's capability to identify anomalies in the breast that may or may not be cancerous. This concept was also extended and validated in an experimental trial; the fabricated device was able to detect a small PEC sphere of 10 mm diameter. Both results, simulation and experimental test, show the potential of this device in successfully detecting breast cancer. This study also explores, in part four, the classification between healthy and unhealthy breast tissue by integrating a microwave probe and a self-organized algorithm. This system collects the reflection coefficients from multiple breast samples where there may or may not be the presence of cancerous tissue. The probe showed remarkable capability to detect the changes in dielectric properties in the breast tissue, and the self-organized algorithm addressed the challenge of classifying this complex data overpassing traditional machine learning methods. We tested the capabilities of this detection system under simulation and experimental validation.
Protection of Multi-Terminal HVDC Grids
(University of Waterloo, 2025-04-28) Radwan, Mohamed
This thesis introduces four novel protection schemes for multi-terminal HVDC grids: primary protection, fault location identification, and breaker/relay failure backup protection schemes. The proposed single-ended primary protection scheme employs Hilbert-Huang Transform (HHT) to extract two instantaneous features from local voltage measurements, namely the instantaneous frequency and energy. Abrupt changes in these features during internal faults are detected as outliers. Simultaneous outliers in the extracted features correspond to internal faults, which offers a setting-less fault detection criterion; thus, eliminating the need for simulation-based and grid-specific thresholds. The proposed double-ended fault location identification scheme employs the setting-less outlier-based criterion to capture the arrival instants of the initial fault-induced backward traveling waves at both line terminals, based on which the fault location is identified with high precision.
A Study of the Eroding Dry-Band Arcing on Silicone Rubber Insulation Using Ultra-High Frequency Detection Technique
(University of Waterloo, 2025-04-25) Mehmood, Basharat
Erosion of silicone rubber housing material is a major cause of failure in outdoor insulators used on overhead power distribution and transmission networks. Dry-band arcing, resulting from leakage currents on the insulator surface, is the main culprit of the erosion of silicone rubber housing material. The inclined plane tracking and erosion test, standardized in IEC 60587 and ASTM D2303, has been a key tool for testing outdoor insulating materials under dry-band arcing at the material development stage, enabling efficient material ranking for field applications. Extensive research on the erosion mechanism of silicone rubber has been carried out in the standard test, and accordingly, the reliable assessment of erosion has facilitated the development of formulations with acceptable performance for outdoor insulation applications. The degree of erosion of silicone rubber varies depending on the severity of dry band arcing during the inclined plane test, with highly severe arcing—referred to as critical or eroding dry-band arcing—often leading to deep erosion. The two inclined plane test standards specify different testing methods, range of test voltages, and failure criteria for the assessment of erosion on silicone rubber. However, the testing methods specified in the standards have often been used interchangeably in different studies, and there is no general agreement on which method is most appropriate for the reliable evaluation of erosion on silicone rubber. In addition, there are wide discrepancies in the selection of test voltages, and no unified agreement exists on the most suitable voltage level for testing silicone rubber materials, leading to inconsistencies in test outcomes. Another crucial aspect of the inclined plane test in evaluating the performance of insulating materials is identifying erosion failure, for which different criteria are specified in the two standards. However, the failure criteria specified in the standards are also applied inconsistently across various studies, with different criteria being used. While some criteria may be effective for certain materials, they are not always applicable to SR—particularly under DC voltages. These inconsistencies in testing methods, voltage selection, and failure criteria largely stem from a limited understanding of the erosion failure mechanisms of silicone rubber, highlighting a critical research gap that needs to be addressed. Additionally, the erosion mechanism of silicone rubber has been extensively studied in the inclined plane test based on leakage current techniques. While these methods correlate leakage current with dry-band arcing severity during the test, the onset of erosion or material failure could not be reliably detected with these techniques. In particular, no significant focus has been given to the erosion failure mechanisms on silicone rubber insulation under the eroding dry-band arcing. Consequently, the underlying physics of the eroding dry-band arcing driving erosion failure of silicone rubber remains inadequately understood, presenting a critical technical gap. This highlights the critical need for a reliable detection method of the eroding dry-band arcing to identify deep erosion failure during the inclined plane test. This thesis provides a mechanistic understanding of the erosion failure mechanisms of silicone rubber particularly through a reliable detection of the eroding dry-band arcing in the inclined plane test. To achieve this, ultra-high frequency detection of the eroding dry-band arcing is introduced as a reliable technique to identify erosion failure of silicone rubber in the AC and DC inclined plane tests. This detection method will serve as a foundation for improving the elucidation of the mechanisms of eroding dry-band arcing that drive erosion failure in silicone rubber. Moreover, it is a critical step towards improving the standards by introducing a clear erosion failure criterion for silicone rubber in the inclined plane test.
Broadband Linear Modulator Driver Design for High-data-rate Wireline Communications
(University of Waterloo, 2025-04-24) Cui, Bolun
This thesis presents the design, analysis, and implementation of a broadband optical modulator driver targeting high-speed wireline communication systems in a 22-nm FD-SOI CMOS technology. The driver is designed to meet stringent specifications for 15.7-dB gain, 90-GHz bandwidth, linear output swing of 4 Vppd, and minimum energy per bit, with a focus on compensating for high-frequency losses in optical modulators through the driver's 9-dB gain peaking. The core of the design revolves around a two-stage broadband amplifier, incorporating a novel shunt/double-series interstage network to achieve amplitude peaking at 75 GHz and extended bandwidth. The interstage network is optimized to provide a 4-pole, 1-zero transfer function, enabling efficient inductive peaking and minimizing in-band ripple. The preamp stage features a PFET cascode amplifier with a shunt/double-series input matching network, designed to achieve broadband input matching up to 80 GHz. The postamp stage employs a parallel-path NFET differential cascode amplifier with an output combining network, optimized for more than 15.2-dBm output 1-dB compression point up to 75 GHz. Experimental results demonstrate the driver's performance across a wide range of operating conditions. The prototype achieves a small-signal bandwidth of over 90 GHz, with a peak gain of 20.4 dB at 76 GHz for a 60-Ohm differential load. The driver demonstrates a low-frequency gain of 15 dB and maintains a flat group delay of 17 ps up to 60 GHz, peaking at 30 ps at 76 GHz. The driver's linearity is characterized using the 1-dB compression point (OP-1dB) and total harmonic distortion (THD). The OP-1dB remains above 13 dBm, equivalent to a differential output swing of 4 Vppd, up to 75 GHz across a 100-Ohm differential load. The THD is measured at 1.6% for a 1-GHz input signal with a 4-Vppd output swing. Time-domain measurements demonstrate the driver's ability to transmit 140-GBaud NRZ signals with a 4-Vppd eye amplitude. For 100-GBaud PAM-4 signals, the driver achieves a 4-Vppd outer optical-modulation amplitude, but the top and bottom eyes begin to close due to the driver's frequency response and distortion with 100-Ohm differential loads. The use of a 4-tap feedforward equalizer (FFE) compensates for frequency response limitations, resulting in improved eye opening. The driver's performance is compared with other state-of-the-art broadband drivers, highlighting its competitive bandwidth, linearity, and energy efficiency. The prototype occupies a core area of 0.17 mm2 and a total area of 0.7 mm2, including pads. The results provide valuable insights into the design and optimization of broadband drivers for high-speed communication systems.
Dual Energy X-ray for Quantitative Analysis of areal Bone Mineral Density (aBMD) using a stacked Flat Panel Detector (FPD)
(University of Waterloo, 2025-04-23) Madani, Nelly
This study explores the application of a stacked flat panel detector (FPD) system for spectral dual-energy X-ray imaging aimed at quantifying areal bone mineral density (aBMD). The proposed technique enables single-exposure dual-energy acquisition using a conventional cone-beam X-ray source, eliminating the need for sequential exposures or specialized hardware. This method facilitates simultaneous high- and low-energy imaging by capturing energy-separated data directly at the detector level, thereby enhancing workflow efficiency and spatial alignment. Experimental validation using bone-equivalent phantoms demonstrates the system's potential for accurate aBMD estimation, with robustness against positional variation and tissue scatter. The findings suggest a promising pathway toward more accessible, portable, and cost-effective bone health assessment in clinical and point-of-care settings.
Resource Allocation for Reconfigurable Intelligent Surface assisted Wireless Communications
(University of Waterloo, 2025-04-22) Abouamer, Mahmoud
Reconfigurable intelligent surfaces (RIS) offer a promising solution to meet the growing demand for connectivity while ensuring stringent quality-of-service (QoS) requirements. An RIS consists of a 2D array of nearly passive reflecting elements, which can be configured to dynamically shape incident electromagnetic waves, thereby engineering the channel between a transmitter and a receiver. An RIS, which typically lacks active components (e.g., RF chains), provides a cost-effective, nearly passive solution for smart radio environments. However, its inability to perform advanced digital signal processing presents challenges, requiring channel state information (CSI) acquisition and RIS configuration optimization to be handled at the transmitter or receiver side, with control signals fed back to the RIS. Thus, efficient channel estimation and optimization of RIS reflection coefficients are essential for practical integration into existing communications systems. Subsequently, this research focuses on addressing these challenges by investigating resource allocation problems in RIS-assisted systems and developing efficient RIS configuration schemes for various practical scenarios. First, we focus on RIS design in multi-user frequency-division-duplexing (FDD) and time-division-duplexing (TDD) systems, proposing a joint uplink-downlink RIS design where the same configuration supports both transmissions. In FDD, this is essential as uplink (UL) and downlink (DL) occur simultaneously, requiring a joint configuration. In TDD, while not strictly necessary, a joint design reduces feedback overhead, power consumption, and configuration periods associated with updating the RIS. To compute the trade-off between uplink and downlink rates achieved by a joint design, a weighted-sum problem is formulated and optimized using a developed block-coordinate descent (BCD) algorithm. The resulting uplink-downlink trade-off regions are investigated by numerical simulation to gain insights into different scenarios. For many considered scenarios, the proposed joint design is shown to bring significant improvements over the fixed-uplink (fixed-downlink) heuristic of using the RIS configuration optimized for uplink(downlink) to assist downlink (uplink) transmissions. Moreover, the proposed joint design substantially bridges the gap to the individual design upper bound of allowing different RIS configurations in uplink and downlink. In the second part, we develop a learning-based framework that directly exploits noisy pilots to optimize a multi-user RIS system while accommodating different service priorities and fairness via user weights. Towards this goal, an adaptive beamforming configuration problem is formulated to generate the RIS system’s beamforming configurations that optimize the weighted sum-rate (WSR). Under mild regularity conditions, this problem is shown to attain a maximum. To learn approximate solutions, a novel hypernetwork-based beamforming (HNB) framework is proposed. Particularly, a beamforming network (BFN) exploits available information, including noisy pilots, to generate optimized beamforming configurations. Rather than learning one BFN, a hypernetwork is trained to dynamically generate BFN learning parameters from an input conditioning vector. When the conditioning vector is chosen as the user weights, the trained HNB can tune the BFN to the user weights without the need for retraining. Numerical experiments demonstrate that tuning allows the proposed HNB to perform close to a block-coordinate descent with perfect CSI benchmark and significantly outperform static learning where a BFN is directly trained to optimize beamforming configurations. Additionally, employing the HNB to also tune the BFN to location information considerably reduces the pilots needed to generate optimized beamforming configurations. In the third part, to enable RIS implementation with diverse technologies, we develop RIS configuration schemes under relaxed hardware constraints. Particularly, we consider the case where an RIS is configured with discrete phase shifts and the amplitude response associated with an RIS element is non-linearly dependent on the phase shift introduced by the element. This RIS configuration problem is formulated as a constrained virtual-channel selection problem. Subsequently, it is demonstrated that the configuration problem, when restricted to subsets of the virtual channels where the maximum phase-variation is θmax < π, can be bounded between monotone upper bound (UB) and lower bound (LB) matroid-constrained problems. We illustrate discrete RIS configuration problems whose optimal solution satisfy this maximum phase-variation property. Moreover, under the maximum phase-variation property, performance guarantees are obtained using a low-complexity configuration scheme. Motivated by this, a practical low-complexity configuration framework is proposed to optimize general discrete RIS problems. Numerical experiments demonstrate the efficacy of the proposed RIS configuration scheme.
Direct Observation of Post-Sonoporation Membrane Resealing in the Giant Unilamellar Vesicle Model
(University of Waterloo, 2025-04-22) Wu, Lyuyuan
Sonoporation is a microbubble (MB)-mediated ultrasonic cavitation, which has attracted much attention in recent years due to its high efficiency and precision in non-invasive drug and gene delivery (Bouakaz et al., 2016; Helfield et al., 2016). However, the cell spontaneous repair process of the membrane after sonoporation, especially the dynamic mechanism of membrane resealing, has not been fully elucidated (Rich et al., 2022). Among them, a key mechanistic question is whether biophysical forces (such as membrane tension) play an important role in the membrane resealing process. However, this issue remains challenging due to the large number of biological processes in cells that may affect the dynamic behavior of membranes, such as cytoskeletal rearrangement (Chen et al., 2014) and endocytosis (Delalande et al., 2015; Zeghimi et al., 2015). To address this issue, this study innovatively used giant unilamellar vesicles (GUVs) composed only of phospholipid bilayers as a minimally simplified membrane research model to explore the membrane reseal mechanism in sonoporation. In the methodology, GUVs were successfully prepared by electroforming. Combined with a coupled ultrasound platform of fluorescence microscopy and high- speed camera, the GUV and microbubble mixed system was exposed in a precise controllable ultrasound field (ultrasonic frequency: 1 MHz, 10% duty cycle, 1Vpp) to simulate the sonoporation. The experimental results showed that GUVs were able to spontaneously reseal their membrane pore after sonoporation, indicating that the phospholipid bilayer has the ability to self-reseal even without the contribution of cell-mediated physiological processes. Further statistical analysis showed that the size of the microbubble had a significant effect on the GUV self reseal results: smaller microbubbles (2–3 μm in diameter) usually formed reversible and easily healed micropores, while larger microbubbles (>8 μm) often led to severe holes that were difficult to bridge (p < 0.01), ultimately destroying the stability of the vesicles. This discovery demonstrated for the first time the key role of biophysical forces in the membrane reseal stage after sonoporation. This study not only provides a new experimental model and platform based on GUVs, laying the foundation for future studies of the biophysical mechanism of sonoporation, but also provides a new perspective for a deeper understanding of the biophysical nature of membrane reseal.
Non-Intrusive Diagnostics of Outdoor Ceramic Insulators Using Ultrasonic Signatures and Deep Learning Models
(University of Waterloo, 2025-04-21) Lutfi, Abdulla
Ceramic insulators have been widely used in overhead power lines for over a century. However, in recent years, transmission and distribution networks have been gradually shifting toward polymeric insulators. Despite this transition, many ceramic insulators remain in service, and a significant portion are now approaching or exceeding their intended lifespans. This aging infrastructure poses an increasing risk of sudden failure, thereby compromising network reliability. Insulator failures account for nearly half of maintenance costs in transmission lines [1], prompting a growing demand among utilities for fast, reliable, and cost-effective condition monitoring systems. Defective ceramic insulators that experience internal punctures, broken discs, or cracks will ultimately initiate partial discharge (PD) activities. Additionally, ceramic insulators exposed to high contamination levels are prone to dry band arcing (DBA), which increases the risk of flashover. Both PD and DBA emit electromagnetic, ultraviolet, infrared and/or ultrasonic radiation, serving as critical indicators that can trigger corrective maintenance actions. Employing sensors to detect these early-stage discharge activities is essential for preventing insulator failure and reducing the risk of power outages. Furthermore, insulator strings often exhibit multiple concurrent defects, resulting in various discharge activities—such as corona, PD, or DBA—each characterized by distinct properties. The overlapping nature of these discharge activities poses significant challenges to accurate diagnosis. This thesis introduces a novel, non-contact method for assessing the condition of outdoor ceramic insulators by employing an ultrasonic sensor in conjunction with deep learning techniques to detect and classify insulator defects. Notably, it demonstrates how the strong directionality of ultrasonic sensors can be leveraged to indirectly identify internal punctures by monitoring surface discharges on adjacent discs, overcoming attenuation limitations caused by the porcelain body and metallic caps. The dissertation is structured into three phases. In the first phase, ultrasonic data from defective insulator strings is generated under controlled laboratory conditions. A multi-class classification model is developed and trained to diagnose individual defects; the model’s performance is then tested in both laboratory and field environments to evaluate its robustness and real-world applicability. The second phase extends this approach to the diagnosis of insulators with multiple concurrent defects. Here, a multi-label classification model is developed to identify and categorize overlapping defect signatures within a single insulator. This approach captures multiple defect types simultaneously, thereby enhancing diagnostic accuracy and reflecting the true operational conditions of outdoor insulators, where different kinds of degradation can co-exist. The final phase involves in-depth analysis of ultrasonic signals by leveraging model-learned features to identify distinct temporal characteristics for each defect type. This enables precise defect characterization and further boosts the accuracy and reliability of outdoor insulator condition assessment. Additionally, findings from Shannon entropy analyses corroborate the presence of unique entropy profiles for different defect classes, improving classification performance—though internal puncture and corona classes can exhibit overlapping energy and entropy characteristics. The results highlight the potential for real-time, non-intrusive monitoring, while emphasizing future work on advanced time-frequency analysis and exploring diagnostic methods for polymeric insulators to address broader asset management challenges.
FPGA-Accelerated Deep Learning for Denoising Low-Dose PET Scans
(University of Waterloo, 2025-04-17) Dao, Eric-Khang Tan
Positron Emission Tomography (PET) is an essential imaging technique used in clinical settings for diagnosing conditions such as cancer and neurological disorders; however, its dependence on radiopharmaceuticals poses potential radiation exposure risks. Lowering the administered dose can help improve patient safety but results in imagery with reduced Signal-to-Noise Ratio (SNR), impacting diagnostic accuracy. The trade-off between minimizing radiation exposure and maintaining image quality remains a key challenge in PET imaging. Recently, deep learning-based denoising techniques, such as Denoising Convolutional Neural Network (DnCNN), have proven effective in restoring noisy images to standard quality. Traditional implementations relying on CPUs and GPUs are often constrained by high power consumption and hardware overhead, limiting feasibility in edge-compute applications. To address these challenges, this thesis explores FPGA-based acceleration for PET image denoising. A dataset is constructed using PET scans from 10 Alzheimer’s disease patients from the ADNI database, with only 0.5% of the original radiotracer dose used. A software-based implementation is developed using a proposed U-Net-like architecture, then ported to an FPGA using OpenVINO and Intel’s FPGA AI Suite for hardware emulation. Experimental results show the FPGA implementation offers a 77% improvement in performance-to-watt ratio compared to the GPU-based solution, and a 2x reduction in latency compared to the CPU-based solution.

Browse

Recent Submissions