26734 research outputs found
Sort by
A Novel Statistical and Experimental Framework for Real-Time Single Cell Classification using Cell Vibrations
No full textCell vibrations are mechanical oscillations generated by living cells which have emerged as a promising biophysical marker reflecting cellular state, function, and identity. Previous studies have identified the presence of cell-specific vibrations; however, their clinical translation has been hindered by methodological inconsistencies and a lack of available analytical frameworks. This thesis presents a novel methodology for non-invasive cellular classification utilizing these vibrations through a technique called Cell Vibrational Profiling (CVP), wherein we acquire data on these nanoscale mechanical oscillations measured via Optical Tweezers (OT). Here we address limitations through three major contributions: first, we advance the real-time vibration acquisition technique through the integration of microinjectors supporting fluidic treatments during vibration experiments, next we perform a comprehensive analysis of experimental factors influencing cell vibrational data, and finally we develop a robust statistical framework called the Vibration Scanner Analysis Suite (VSAS). Using U251 glioblastoma cells as a primary model, significant vibrations were identified at 402.6, 1254.6, 1909.0, 2169.4, and 3462.8 Hz, which were statistically different from polystyrene bead controls (p<0.0001). Using a fixative (Paraformaldehyde) we observed a three-fold decrease in the Coefficient of Variance (COV) produced by U251 cells, confirming the metabolic origin of these signals. Next, we validate this technique by reanalyzing existing datasets and compare experimental conditions to find those that produce maximum differentiation. This was done using Partial Least-Squares Discriminant Analysis (PLS-DA). We found that differentiation was maximized for A549 alveolar epithelium cells when data was collected in conventional petri-dishes (F1=0.89) and while their cell-cycle was synchronized in G0/G1 using serum-deprivation (F1 = 0.82). All findings were generated through the VSAS, which is an automated statistical analysis program purposefully designed to address heterogeneity and noise across CVP datasets and present objective, data-driven results. Together, this research combines both experimental and analytical advancements and establishes a comprehensive pipeline for cell vibration studies with potential applications in cancer detection, pharmaceutical testing, and realtime diagnostics in surgical settings
Instrumentation, multi-omics, and computational approaches to elucidate cell state dynamics in human induced pluripotent stem cell expansion biomanufacturing
No full textExpansion bioprocessing enables the scalable production of human induced pluripotent stem cells (hiPSCs) for therapeutic use, generating populations in the millions to trillions. Despite growing global infrastructure and increasing clinical trial activity, metrics to assess cell state dynamics governing pluripotency in the process remain insufficient. This raises the question of the role reinforcement of pluripotent phenotype during expansion plays in observation of suspended or failed trials due to safety or efficacy concerns. While cell therapy holds curative potential, fully unlocking its promise remains a challenge. This dissertation addresses key barriers by studying how bioprocess conditions in the artificial niche influence hiPSC state dynamics during expansion. The central hypothesis is that modeling and optimizing control over the networked determinants of cell state in response to process variables is critical to facilitating optimal therapeutic derivation. Four aims guided this work. First, the influence of oxygenation on extracellular metabolite dynamics was investigated. Distinct profiles of metabolite depletion and accumulation were observed over time, implicating metabolic pathways in cell state adaptation. These findings suggested that monitoring parameters like dissolved oxygen, pH, and cell density might predict surrogate metrics associated with metabolic phenotype. The second aim addressed limitations in biomanufacturing infrastructure by optimizing a scale-down culture platform. This system enabled modulation of process parameters to study hiPSC dynamics under conditions mimicking larger-scale processes. Third, a non-invasive optoelectronic system was developed to monitor oxygen consumption and extracellular acidification in real time. The system reliably captured dynamic changes in response to oxygenation, supporting its use in biomanufacturing studies. The fourth aim integrated multi-omics analysis to assess transcriptional and metabolic remodeling under varying oxygen and agitation conditions. Results revealed rapid, condition-dependent shifts in intracellular networks that could not be fully captured by surrogate measurements alone. These insights underscore the need for higher-resolution, time-resolved data to model and control cell state effectively. Together, this work provides a foundation for integrating real-time monitoring and computational modeling to enhance control over hiPSC cell state in vitro, advancing the field toward more predictable and effective therapeutic applications
For Your Eyes Only: A Look into the Legal Feasibility of Exceptional Access to End-to-End Encrypted Messages
No full textModern messaging applications have revolutionized the world of interpersonal communications. They allow us to seamlessly communicate with people across the globe. These platforms, whose infrastructure operates on the internet, secure their communications through cryptography with most of them providing end-to-end encryption. This means that no one except the participants of the communications can access these messages, not even the platforms themselves. This poses a problem for law enforcement who then face hindrances in accessing those messages for investigative purposes. As a result, law enforcement agencies and governments across the world have been calling on these platforms to provide ways to facilitate lawful exceptional access to these communications. While the intentions behind lawful exceptional access may be noble, facilitating such mechanisms in end-to-end encrypted communications can create more problems than it seeks
to solve. I will assess some of these legal problems in my thesis. More specifically, I will look at the economic, constitutional, and evidentiary concerns posed by facilitating lawful exceptional access to end-to-end encrypted messages. By understanding the basics of cryptography and how it is used in end-to-end encrypted messaging, we will see how messaging platforms may be inclined to leave Canada instead of complying with the exceptional access mandate. By assessing these technological mechanisms against the Canadian constitutional law, we will see that facilitating lawful exceptional access may be unconstitutional as violating the right against unreasonable search and seizure guaranteed by the Section 8 of the Canadian Charter of Rights and Freedoms. We will also see that exceptional access will compromise the integrity and authentication provided by end-to-end encrypted messages which may further weaken the evidentiary value accorded to these messages. The takeaway from this thesis is that introducing exceptional access mechanisms in end-to-end encrypted messaging would go against the interests of law enforcement. To ensure that cryptography does not impede police investigations, law enforcement should focus on alternative investigative techniques which do not compromise the embedded security mechanisms of these messaging platforms, which have become ubiquitous across the world
Thermochemical conversion of high-density polyethylene (HDPE) to value-added products using cobalt-based catalysts
No full textEmerging global plastic demand has escalated the generation of waste plastic, raising serious concerns for the environment and society. High-density polyethylene (HDPE) covers 13% of the total waste plastic. Pyrolysis is a thermochemical method of treating plastic thermally under an inert atmosphere to convert HDPE into valuable products, such as wax, oil, gas, and char. The current thesis explores the utilization of cobalt-based catalysts in the HDPE pyrolysis. The study investigates the effects of cobalt loadings in Co3O4/Al2O3 catalysts and reveals that an increase in cobalt loading resulted in gas production, while lowering the wax formation. The cobalt phase plays a crucial role in determining product selectivity and catalyst stability during the pyrolysis process. The Co3O4 catalyst favours gas production but is an unstable catalyst, whereas CoAl2O4 promotes wax and is a stable catalyst. The results highlight the potential of cobalt-based catalysts in converting HDPE into valuable products and contributing to plastic waste management
Parsing The Electrophysiological Signatures of Spatially Selective Versus Distributed Attention in Neuronal Population Data
No full textAttention efficiently allocates our limited cognitive resources by prioritizing specific sensory inputs for further processing to enhance performance and speed reaction times. These behavioural changes are coupled to concurrent changes in neurophysiological processes such as enhanced responses of visual neurons, reduced influence of distracting information, and increased arousal. While these processes covary, emerging literature shows that they arise from distinct neurobiological processes. Given this, parsing the diverse neural mechanisms that support attention is critical to develop new diagnostic tools and therapeutics for attention-related disorders, ultimately improving mental health outcomes. While attentional states such as selective and distributed attention enhance behavioural performance, emerging evidence suggests these effects are driven by dissociable neural mechanisms. In this thesis, we investigate how attentional state modulates neural population dynamics in visual area V4 using a large dataset of 1,194 neurons recorded from two macaque monkeys during a visual orientation change detection task. By contrasting selective and distributed attention conditions, we examine how shifts in attentional allocations reshape correlations, gain modulation, and normalization processes among neurons. Using machine learning techniques including tensor decomposition, Isomap, and subspace projections we identify low-dimensional neural signatures that differentiate attentional states. Our findings show that distinct geometries of neural population activity reflect attentional allocation, independent of stimulus identity, and are shaped by spatially tuned normalization mechanisms. These results deepen our understanding of how cognitive state reshapes sensory processing and may inform future approaches for diagnosing and treating attention-related disorders as well as providing new insights into the neuronal mechanisms that support visual attention to optimize behaviour
Towards Developing a Foundation Model for Agricultural Vision
No full textThe rapid advancements in artificial intelligence and computer vision have enabled the development of powerful models for natural image understanding. However, adapting these advances to domain-specific tasks in agriculture—characterized by limited annotated data, high visual complexity, and domain shift—remains an open challenge. This thesis aims to address this gap by developing two foundation-model-inspired frameworks tailored for precision agriculture: WheatSAM and Agri-FM+. WheatSAM presents an adaptation of the Segment Anything Model (SAM) for instance-level wheat head segmentation. It introduces a two-stage segmentation framework combining automated prompt generation via object detection with a novel error-aware bounding box perturbation strategy to enhance SAM’s robustness to noisy prompts. WheatSAM exhibits consistent improvements over baseline SAM models (ViT-B/L/H) and traditional segmentation pipelines, achieving high segmentation fidelity in complex field conditions. Complementing this, Agri-FM+ is a self-supervised foundation model pretrained on a curated dataset of 147,000 close-range agricultural images using contrastive learning and continual adaptation from ImageNet. Designed to learn generalizable visual representations for crop phenotyping, Agri-FM+ demonstrates superior performance across eight downstream agricultural tasks, including object detection, semantic segmentation, and instance segmentation. Notably, it outperforms both ImageNet-supervised and from-scratch-trained baselines, particularly under limited supervision, highlighting the value of domain-adapted pretraining. Together, these contributions bridge recent advances in foundation models with the specific demands of agricultural vision, offering scalable, reusable, and label-efficient solutions for real-world phenotyping applications. This thesis establishes a new paradigm for domain-specialized foundation models in agriculture, laying the groundwork for future research in multimodal, federated, and open-world agricultural AI systems
Exploring Career Development with Young People Who Age Out of Out-of-Home Care in Canada: What Helps and Hinders in the Pursuit of Career Success?
No full textThe three manuscripts that comprise this program of research expanded the understanding of career development experiences among young people from out-of-home care and identified ways in which career success may be promoted for this population. The first manuscript is a scoping review of literature on the employment experiences and outcomes of young people from out-of-home care. This review highlighted the various intrapersonal, interpersonal, and systemic challenges that young people from out-of-home care encounter in preparing to connect to the labor market and secure gainful employment and presents some of the few factors that are known to support these processes for this population. The other two
manuscripts were created from a data-driven study that explored what helped, hindered, and was missed in the career development experiences of a sample of young people who aged out of out-of-home care and self-identified achieving career success. Specifically, the second manuscript identifies categories of factors that created challenges and barriers, as well as categories that facilitated and promoted, career development and career success with this population. The third manuscript provides recommendations for practice and policy to support the career development of this population, based on categories of factors that participants did not experience but believed would have been helpful in their journeys of career success. Finally, important findings across all three manuscripts are integrated to discuss the overall implications of this research for career development, child welfare, and counselling psychology
An Optimization Framework for Next-Generation Video Streaming Systems
No full textThe rapid expansion of internet connectivity and digital media consumption has firmly established video streaming as the dominant contributor to global internet traffic. With the advent of immersive formats such as 360-degree and volumetric video, user experiences are shifting from passive viewing to interactive, spatially-aware engagement. However, these advanced formats introduce significant challenges, including extremely high bandwidth demands, real-time processing constraints, viewport prediction inaccuracies, and unpredictable user behavior in three and six degrees of freedom (3DoF and 6DoF) environments. The growing adoption of such technologies has placed considerable strain on existing network infrastructures and streaming platforms, exposing critical limitations in scalability, responsiveness, and quality assurance. Despite temporary mitigations—such as lowering streaming resolution—users often attribute service degradation to Internet Service Providers (ISPs), underscoring the urgent need to better balance Quality of Service (QoS) and Quality of Experience (QoE) in bandwidth-constrained and dynamic environments. This thesis presents a dual-layer optimization framework to address these challenges at both the application and network layers. At the application layer, two frameworks—Opt360 and OptVV—introduce real-time optimization models tailored to immersive media. Opt360 enforces stall-free playback as a strict constraint, accounting for dynamic viewport changes and prediction inaccuracies. It also introduces a staged streaming enhancement to the DASH protocol to improve the quality of field-of-view (FoV) content. Experimental results show that Opt360 ensures seamless playback, remains resilient to head movement variations, and maintains high video quality across fluctuating bandwidth conditions. OptVV extends this approach to volumetric video streaming, offering a comprehensive optimization model built around three components: (1) adaptive QoE optimization to balance video quality and resource efficiency in real time; (2) a DASH-based scheduling strategy tailored to the unique demands of volumetric content; and (3) resource-aware decoding optimization for managing computational complexity. Evaluations reveal that OptVV achieves up to 80% bandwidth savings, 177% improvement in viewport quality, 51\% fewer playback stalls, and 72% faster decoding performance compared to existing solutions. At the network layer, this thesis introduces SOFT-Stream, a scalable, SDN-enabled optimization framework for managing DASH sessions in large-scale ISP environments. SOFT-Stream jointly optimizes the quantity and quality of concurrent streaming sessions, while addressing practical concerns such as dynamic session behavior and real-time optimization overhead. Experimental validation shows that SOFT-Stream outperforms conventional network resource management schemes, delivering a 52% increase in session acceptance, a 200% improvement in bandwidth allocation, a 70% reduction in bandwidth wastage, and a 77% speed-up in computation time. The framework also delivers consistently smoother playback and enhanced end-user experiences across varying network conditions. Together, the contributions of this thesis offer a robust, end-to-end solution for next-generation immersive video delivery
Federated Pseudo-labeling: A Data-Centric, Privacy-Preserving Framework for Medical Image Segmentation
No full textMedical image segmentation is essential for clinical diagnosis, treatment planning, and dis-ease monitoring. However, its advancement is impeded by strict data privacy regulations, the high cost of expert annotations, and significant variation in imaging protocols across institutions. These factors restrict the generation of large, centralized annotated datasets, limiting the generalizability of traditional deep learning models. Federated Learning (FL) has emerged as a decentralized alternative, enabling model training without data sharing. However, existing FL methods often rely on iterative parameter sharing and require uniform model architectures, which limit flexibility between institutions with diverse computational infrastructures and datasets. Parameter share also introduces residual privacy risks and significant communication overhead. Performance of FL degrades under non-identically dis-tributed data, a common characteristic in medical imaging. To address these challenges, we propose DCFed, a novel data-centric, semi-supervised FL framework that eliminates parameter sharing entirely. DCFed leverages federated pseudo-labeling on publicly available unlabeled datasets, using logit-based aggregation and uncer-tainty estimation. Clients independently train on their private datasets while collaboratively generating and refining pseudo-labels on a shared public dataset. These pseudo-labels are redistributed to improve local models without transmitting any weights or exposing sensi-tive data. Each client utilizes a customized U-Net backbone, enhanced with residual blocks, atrous spatial pyramid pooling for multi-scale feature extraction, and convolutional block attention modules to refine spatial and channel-wise representations. We have evaluated DCFed on two clinically relevant tasks: breast cancer segmentation from ultrasound images and skin lesion segmentation from dermoscopic images and observe performance improvements of up to 8.9% and 3.7%, respectively, over strong local baselines. Compared to standard FL, DCFed reduces communication overhead by more than 215× while achieving higher segmentation accuracy than both centralized and conventional FL approaches. Moreover, the performance of the proposed approach has been compared with standard FL methods such as FedAvg, FedNova, FedOpt and FedProx and DCFed outper-forms all these methods across most of the clients. Notably, DCFed is scalable across diverse client model architectures, accommodates clients with varying data volumes and labeling availability, and supports flexible collaboration settings, making it a highly practical and privacy-preserving solution for real-world medical image segmentation
Morphology, Phylogeny, and the Evolution of Eutyrannosauria (Dinosauria: Tyrannosauroidea)
No full textTyrannosauroidea were a group of carnivorous theropod dinosaurs that are known from the Late Jurassic and Cretaceous of Eurasia and North America. Despite over a century of paleontological discoveries, our knowledge of tyrannosauroid paleobiology is still lacking. Here, anatomical examination of specimens and analytical techniques are used to further understand or clarify the morphology, ontogeny, taxonomy and phylogeny of tyrannosauroids and shed new light on the
evolutionary history of eutyrannosaurian tyrannosauroids, the archetypal apex predatory subclade of Tyrannosauroidea that includes the iconic Tyrannosaurus rex. Detailed anatomical descriptions are provided for three species of tyrannosauroid, Gorgosaurus libratus, Thanatotheristes degrootorum, and Khankhuuluu mongoliensis. A detailed description of the ontogenetic changes in the braincase is provided for the well-known species Gorgosaurus libratus, whereas a complete anatomical description is provided for specimens of the latter two recently described species. A new phylogenetic character matrix was constructed using these descriptions and comparative anatomical research on many specimens of other tyrannosauroid species. A combination of parsimony phylogenetic analysis, three-rate calibrated time calibration, and ancestral state estimations are implemented to provide new hypotheses for the evolutionary history of Eutyrannosauria and other Tyrannosauroidea. The results show that Khankhuuluu is recovered as the latest diverging non-eutyrannosaurian tyrannosauroid. Eutyrannosaurian origins are rooted in a dispersal of non-eutyrannosaurian tyrannosauroids, like Khankhuuluu, from Asia into North America approximately 91-86 million years ago. General morphological similarities between Khankhuuluu and juvenile eutyrannosaurians provide evidence for peramorphic heterochrony (extended/accelerated growth) as a key factor in the evolution of robust skeletal adaptations that characterize Eutyrannosauria. Eutyrannosauria remained exclusively North American until a dispersal of the clade of Tyrannosaurinae, likely similar to Thanatotheristes, back into Asia approximately 79-78 million years ago. A cladogenic event resulted in a split of this Tyrannosaurinae lineage which gave rise to the gracile, shallow-snouted Alioramini and the massive Tyrannosaurini in Asia; the Tyrannosaurini lineage then dispersed to North America between 73-66 million years ago, ultimately giving rise to Tyrannosaurus rex. While Alioramini and Tyrannosaurini are shown here to share many synapomorphies, other general morphological differences indicate that Alioramini was influenced by paedomorphosis whereas tyrannosaurins, like most other eutyrannosaurians, were influenced by peramorphosis. Stark differences in general morphology may have allowed these two clades to evolve sympatrically in Asia where alioramins fulfilled the mesopredatory role and tyrannosaurins the apex predator role