Abstract

Early detection and effective management of medical conditions significantly enhance patient outcomes, yet challenges persist due to limited resources, particularly in achieving widespread medical expertise. While supervised deep learning excels in medical image segmentation, the scarcity of labeled data poses significant hurdles for clinical applications. In response, self-supervised learning (SSL) systems have gained traction, leveraging unlabeled medical images to extract essential properties. Among SSL methods, contrastive learning stands out, aiming to minimize the distance between similar images while maximizing it between dissimilar ones, thereby enhancing segmentation accuracy. Current SSL methods exhibit limitations, prompting the advancement of unsupervised learning techniques that leverage data’s inherent structure to learn representations without the need for labeled data. A considerable challenge with unsupervised pre-training is domain shift, which occurs when the distribution of data used for pre-training differs from that used for fine-tuning. Addressing a crucial challenge in thoracic disease diagnosis using X-ray imaging, our work in the second chapter explores the need for anatomically accurate findings and why they shouldn’t be overlooked in interpretability assessments. We propose an innovative self-supervised and weakly supervised pre-training pipeline paired with an auxiliary loss followed by supervised fine-tuning to tackle these limitations effectively. Leveraging the Chest X-ray14 dataset for pre-training and the CheXpert dataset for fine-tuning ensures model stability and generalizability across diverse data sources. Specifically, our approach demonstrates a remarkable 31% improvement in Intersection over Union on the NIH CXR dataset, enabling precise classification across 14 chest X-ray categories. While our focus lies on chest X-rays, our proposed approach extends its applicability to various imaging modalities and clinical workflows, showcasing its broad impact in medical imaging and diagnostics. In our third chapter, we expand the Swin Transformer architecture to learn from multiple medical imaging modalities, enhancing downstream performance. Our novel Swin-FUSE (Swin Multi-Modal Fusion for UnSupervised Enhancement) framework offers several advantages: utilizing CTs and MRIs for pre-training to create complementary feature representations, incorporating a domain-invariance module (DIM) to enhance domain adaptability, and demonstrating high generalizability beyond pretrained tasks. Empirical results on publicly available 3D segmentation datasets showcase a marginal 1-2% performance trade-off compared to single-modality models while achieving a remarkable 27% improvement on out-of-distribution modalities. In our fourth chapter, we introduce an attention-based model for the 2023 ISBI X-ray Projectomic Reconstruction Challenge, aimed at fostering exploration into XNH imaging challenges, particularly focusing on the white matter region of the rodent brain. Our proposed AGU-Net model demonstrates enhanced performance in merging segmented regions and producing accurate segmentations with reduced error rates, as evidenced by its higher scores in the Normalized ERL and Rand merge metrics. Our model achieved an outstanding XPRESS score of 0.8159 on the test set, securing 3 rd place. In summary, this thesis presents innovative solutions leveraging self-supervised pre-training and contrastive learning to enhance representation learning in medical imaging. These advancements extend to domain adaptation tasks, ultimately enhancing the interpretability of conventional black-box machine learning models, thus paving the way for more reliable and clinically relevant diagnostic tools

Abstract

Systems of linear equations are fundamental to numerous scientific and engineering domains, often arising from the discretization of partial differential equations (PDEs) used to model real-world phenomena. These systems can be computationally expensive to solve, especially for large-scale problems. Therefore, developing efficient solvers for linear systems is crucial for advancing scientific computing. This thesis examines the use of recursive partitioning for solving linear systems, focusing on systems derived from 3-dimensional structured grids. The research emphasizes the development of a preconditioner based on this partitioning strategy to enhance the efficiency of iterative solvers, notably the Preconditioned Conjugate Gradient (PCG) method. By decomposing the original problem into smaller, more manageable sub-problems, this approach aims to reduce the overall computational time. This strategy aligns with the principles of parallel computing, where tasks are divided and processed concurrently across multiple processors to achieve faster solutions. The effectiveness of the proposed preconditioner is assessed using Skyscraper problems with checkerboard variable jumps as test cases. These problems, known for their intricate coefficient structures, serve as relevant benchmarks for evaluating the performance gains offered by the partitioning approach.

Abstract

Trajectory planning plays a critical role in autonomous navigation, where the emphasis lies on planning smooth, collision-free trajectories that comply with the robot’s kinematic constraints. It finds itself in various applications, such as autonomous driving, search and rescue, construction, and forestry. In the case of autonomous driving, the idea is to mimic human driving, which involves complex decisions like merging and overtaking, along with lower-level commands. However, standard approaches fail to capture the multi-modal characteristic of human driving. Although autonomous driving finds itself in structured, smooth, planar roads, applications like search and rescue and construction involve uneven terrains. Thus, there should be a distinction in the way trajectory optimization is approached in these scenarios, away from 2D approaches. The first study introduces two methods, which accomplish the task of accomplishing a given high-level driving task. The first method achieves the task of generating trajectories for multiple goals, by parallelizing existing state-of-the-art solvers using multiple CPU threads. The second method introduces a Multi-Convex MPC is based on reformulating the collision avoidance and kinematic constraints and is solved using the Alternating Minimization(AM) technique. The second study introduces an approach to planning trajectories on uneven terrains. It leverages the digital terrain information, represented as a point cloud. The pose of the robot is represented in 6dof. The wheel terrain interaction and pose are expressed as a non-linear problem. The resulting pose prediction, along with Sampling based Model Predictive Control(SMPC), are used to sample smooth and stable trajectories. These studies collectively help advance the understanding of trajectory planning and its implementation in different scenarios.

Abstract

Solar UV radiation challenges the integrity of the genome, causing severe mutations and potentially leading to cancer. Regulatory regions of DNA are particularly important for transcription, and hence preventing damage formation in those regions is essential. When UV damage occurs, it triggers various response systems in the body, the most common being nucleotide excision repair (NER). Global genome nucleotide excision repair (GGNER) takes a broad approach to the repair process, with proteins like Xeroderma Pigmentosum C (XPC) scanning the entire genome for lesions. The mechanism by which XPC identify damage sites from a long, perfectly matched DNA duplex is still not completely resolved. In this thesis, we examine three stages of handling DNA damage: (A) prevention, (B) recognition, and (C) binding to repair protein XPC. Experimentally, it has been found that histone methyltransferase ASH1L prevents the formation of cyclobutane pyrimidine dimers (CPD) in enhancer regions of DNA. Our simulations show that ASH1L binding distorts the geometry of adjacent base pairs, disfavoring the induction of CPD lesions. Subsequently, we studied how the induction of mismatch damage affects the dynamics of DNA fluctuations. FRET lifetime and efficiency analysis has shown that mismatches well recognized by XPC exhibit different intrinsic fluctuations compared to poorly recognized mismatches. We computationally calculated the FRET efficiency for different mismatches from MD simulations. Further, umbrella sampling simulations showed that the fluctuations in the CCC mismatch (a wellrepaired mismatch) can be mapped to a metastable state not accessible to matched DNA or poorly recognized lesions. This suggests a conformational selection mechanism of repair by XPC. Lastly, we investigated the mechanism by which XPC binds to sites of damage. For this study, we used the crystal structure of RAD4 (the yeast ortholog of XPC) bound to pyrimidine (6-4) pyrimidone photoproducts (64pp) to determine the mechanism of RAD4 binding. Using the nudged elastic band method, we determined the minimum energy pathway between the unbound and productively bound states of RAD4. The obtained pathway reveals that the β hairpins of RAD4 cause the DNA to sample multiple twist states, finally reaching a stable twist angle that is favorable for binding. Furthermore, using enhanced sampling in conjunction with our minimum energy path, we determined the energetic cost of some of the key binding events.

Abstract

Inverse Rendering is a core Computer Vision problem as it involves complete decomposition of an image into its constituting atomic components. These components can be stand-alone analyzed or suitably modified and recombined to solve the required image analysis task or achieve the required generative content. Rather than aiming for full decomposition, many applications only require decomposition into only a few factors which themselves are simple combinations of the underlying atomic components. This makes image factorization a critical first step in several computer vision and image processing applications. This factorization could either be optically motivated like reflectance-shading decomposition, white-balancing, illumination spectra-separation etc. or semantically motivated like style-content disentanglement, foreground-background matting etc. In this thesis, we focus on the former and present several image factorization solutions with an aim to use it for a downstream image based rendering application. Initially, we assume Lambertian reflection only under the classical image formation model inspired from the Retinex theory. Our first solution in this category requires multiple images of the scene as input, which we then relax for our second solution which works on the single image input. Afterwards, we propose a novel image formation model based on the specularity of the image content and provide two solutions using the low light enhancement problem as the vehicle for empirical validation. Towards the end, a novel prior induction technique is also presented based on learnable concepts and its utility is shown by improving results of pre-existing state-of-the-art image decomposition networks. We conclude with a summary, limitations, future research directions and possible additional applications. The thesis is organized into four units respectively discussing the problem definition and significance; Lambertian reflection based Intrinsic Image Decomposition problem, specularity respecting novel illumination factorization methods and finally concept based model analysis and conclusion. We hope that with the problems and solutions discussed in this thesis we will be able to define and highlight the importance of image factorization step in multiple vision tasks and pique reader’s interest in this research problem for image generation and beyond.

Abstract

Random Number Generators (RNG) serve as vital tools in various fields, particularly in cryptography and security applications, where the generation of high-quality random bitstreams is paramount. RNGs are essential for ensuring the unpredictability and security of cryptographic protocols and algorithms. They find applications in simulations, gaming, statistical sampling, and numerous other areas where randomness is required. In cryptography, RNGs play a critical role in generating cryptographic keys, initialization vectors, and nonces for secure communication and data encryption. However, the quality and randomness of generated numbers are crucial factors that directly impact the security of cryptographic systems. Traditional RNGs often rely on algorithms or physical sources of randomness, such as thermal noise or radioactive decay. In this thesis, we introduce a fresh method for generating such random bitstreams by harnessing the inherent noise characteristics of ring oscillators. Our study involved the implementation of ring oscillators with varying configurations (3, 5, and 7 stages), diverse geometries, and different startup voltages using Cadence. We measured their total output power, incorporating cumulative noise effects, and then exported these measurements to MATLAB for further analysis. By employing Fast Fourier Transform (FFT)-based techniques, we extracted the overall noise characteristics for each ring oscillator. Utilizing this noise data, we produced separate random bitstreams consisting of 10 million bits for each of the 3-stage, 5-stage, and 7-stage ring oscillators. The final random bitstream, also comprising 10 million bits, was generated by performing bitwise XOR operations on the bitstreams produced by each ring oscillator. The randomness of these generated bitstreams was evaluated using the NIST 800-22 statistical test suite. Impressively, the resultant random bitstream displayed robustness and suitability for cryptographic applications, demonstrating the efficacy of our innovative approach. Leveraging the noise properties of ring oscillators, our method presents a reliable means of generating random bitstreams, with potential applications in secure communications and cryptography. These findings underscore the viability of employing ring oscillators as noise sources for random bit generation, highlighting their effectiveness in meeting stringent randomness criteria.

Abstract

In recent times, the popularity of gym culture has witnessed a remarkable surge. The pursuit of a healthier lifestyle, increased awareness of the benefits of regular exercise, and the desire for physical well-being have contributed to the widespread embrace of gym activities. Individuals across diverse age groups and backgrounds are increasingly recognizing the importance of incorporating fitness into their daily routines, leading to a higher interest in gym memberships and fitness classes. One crucial aspect of effective workouts is the importance of exercise form. Whether engaging in weightlifting, cardio exercises, flexibility routines, or even yoga, the proper execution of movements is crucial for maximizing benefits and preventing injuries. Maintaining correct posture, alignment, and technique not only enhances the efficiency of the workout but also safeguards against potential strains and stresses on the body. Recognizing the significance of exercise form, there is a growing demand for innovative solutions that can assist individuals in ensuring they perform exercises correctly. The need for a device that checks and monitors one�s form has become evident in the fitness landscape. This thesis presents a suit that provides real-time feedback, offering guidance on posture, movement range, and overall form during exercises. By integrating this technology into the fitness routine, individuals can optimize their workouts, reduce the risk of injuries, and enhance the overall effectiveness of their training sessions. This innovative suit employs pressure sensors to precisely calculate the flexion of individual muscle groups, thereby assessing the form of exercise. Utilizing piezo-resistive material, namely velostat, for our pressure sensors ensures dynamic responsiveness. The resistance varies proportionally based on the degree of muscle stretching, enabling the identification of over-stretching or under-stretching that may indicate suboptimal form. By comprehensively analyzing data from all sensors, we can accurately evaluate the proficiency of the exercise form. The focus of this thesis centers on a specialized suit designed to analyze the form of the back muscles during various exercises such as squats, deadlifts, and rows. The suit incorporates multiple sensors strategically placed on the back, seamlessly connected to the ESP32 microcontroller. This setup enables real-time feedback, with all necessary calculations performed on the microcontroller itself. Notably, the suit prioritizes user comfort by offering flexibility and durability without any rigid components that might impede the workout experience. Additionally, the suit features a detachable battery, ensuring its complete washability. Impressively, our suit demonstrates an accuracy range of 72% to 90% across the three specified exercises, highlighting its efficacy in precisely evaluating exercise form. This research not only contributes to the growing field of wearable technology in fitness but also underscores the potential for real-time, personalized feedback to enhance users� workout experiences.

Abstract

In the era of digital intimacy, online dating applications like Bumble have significantly reshaped social interactions and romantic engagements, particularly in India. This thesis delves into the gender disparities manifested within these platforms, focusing on the algorithmic biases present in Bumble�s match recommendations. By examining how these algorithms might favour specific user demographics over others, this research addresses the broader implications of such biases on the visibility and representation of non-binary and marginalized genders. The thesis posits that while these platforms claim neutrality, their underlying algorithms often replicate societal biases, thereby influencing user experiences based on gender. This investigation is anchored in India�s rapidly evolving digital landscape, where traditional norms meet a burgeoning digital dating culture. Employing a mixed-methods approach, this study combines quantitative data analysis with qualitative interviews to offer a comprehensive view of how users perceive and experience gender biases. The quantitative analysis assesses the disparities in match recommendations. At the same time, qualitative insights are gathered from user narratives to explore the subjective impact of these algorithms on individual self-perception and interpersonal interactions. This methodological framework highlights the prevalence of algorithmic bias and enriches the understanding of its practical effects on users. The research findings indicate a significant difference in the algorithmic treatment of users based on gender, revealing a complex interplay of technology and social identity that often reinforces traditional gender roles and stereotypes. The conclusions drawn from this study underscore the urgent need for algorithmic transparency and the integration of ethical considerations in the design and deployment of dating applications. By documenting the nuanced ways in which gender disparities are embedded within digital platforms, this thesis contributes to the critical discourse on digital ethics and algorithmic fairness. The research advocates for more inclusive and equitable design practices that consider the diverse spectrum of gender identities, ultimately aiming to inform developers, policymakers, and the broader community about the challenges and opportunities for fostering inclusivity in digital spaces. The findings pave the way for future research into the inclusivity of digital platforms and serve as a call to action for adopting more responsible technology practices in online dating.

Abstract

The network of cameras around urban cities is increasing in a manner unmanageable by current infrastructure. Video feeds in these networks requires deploying extensive Computer Vision pipelines to generate useful insights. However, poorly designed architectures lead to overspending on the infrastructure by local authorities. This dissertation proposes a scalable distributed video analytics framework that can process thousands of video streams from sources such as CCTV cameras using semantic scene analysis. The main idea is to deploy deep learning pipelines on the fog nodes and generate semantic scene description records (SDRs) of video feeds from the associated CCTV cameras. These SDRs are transmitted to the cloud instead of video frames saving on network bandwidth. Using these SDRs stored on the cloud database, we can answer many complex queries and perform rich video analytics, within extremely low latencies. There is no need to scan and process the video streams again on a per query basis. The software architecture on the fog nodes allows for integrating new deep learning pipelines dynamically into the existing system, thereby supporting novel analytics and queries. We demonstrate the effectiveness of the system by proposing a novel distributed algorithm for real-time vehicle pursuit. The proposed algorithm involves asking multiple spatio-temporal queries in an adaptive fashion to reduce the query processing time and is robust to inaccuracies in the deployed deep learning pipelines and camera failures.

Abstract

Smart Cities represent the nexus of innovation in urban planning, leveraging the Internet of Things (IoT) to foster enhanced urban living. However, the diversity of IoT devices and protocols often creates interoperability barriers. oneM2M, a global standard, addresses this challenge by providing a common service layer that enables seamless communication and data exchange between heterogeneous IoT devices along with features for device management, security, and semantic interoperability. Its resourcebased model and support for various communication protocols make it a key enabler for interoperable smart city solutions. This enables a wide range of intelligent smart city applications, including coordinated traffic management, intelligent energy grids, and responsive public services. This thesis zeroes in on the pivotal aspect of interoperability within Smart Cities, with a particular focus on oneM2M-based interoperability systems. Through a detailed exploration of oneM2M standards and their application in the Smart City context, this research underscores the importance of a unified, standards-based approach to IoT integration, highlighting the role of interoperability in unlocking the full potential of Smart City initiatives. In addressing the critical challenge of enhancing latency and performance in oneM2M-based systems within Smart City frameworks, this thesis recognizes the limitations of existing architectures, which often impede scalability and efficiency. A novel distributed, multi-layered data platform architecture for oneM2M based interoperability platforms is proposed, aiming to demonstrate significant improvements in latency and overall system performance while ensuring scalability across diverse and large-scale Smart City deployments. The research encompasses a comprehensive performance analysis of this architecture versus a centralized alternative within a real-world Smart City Living Lab deployment. Additionally, an exploratory analysis of open-source oneM2M-based interoperability systems (Mobius, OM2M, ACME) using this large-scale system offers insights into architectural choices and their suitability for smart city implementations. Another challenge that becomes fairly evident is the complexity that arises in the usage of the said interoperability solutions based on oneM2M. To increase the ease of use and of the oneM2M based solutions within the IoT ecosystems, this thesis presents the City IoT Operating Platform (ctOP). This platform, meticulously engineered as a lightweight oneM2M wrapper, is designed to enhance interoperability, streamline IoT device integration, and provide robust functionality such as user management and security within real-world smart city deployments. This contribution is pivotal for the IoT field, challenging the status quo of existing oneM2M system design paradigms and proposing a more adaptable, scalable, and user-friendly alternative. Through rigorous analysis, real-world deployment scenarios, and the development of the ctOP platform, this research contributes to both the academic discourse on Smart City development and provides actionable insights for practitioners and policymakers seeking to implement or refine IoT interoperability solutions in urban settings. The ultimate goal is to pave the way for more interconnected, sustainable, and resilient Smart Cities, where technology serves as a backbone for addressing complex urban challenges.