LEE- Hesaplamalı Bilim ve Mühendislik-Yüksek Lisans

Bu koleksiyon için kalıcı URI

http://hdl.handle.net/11527/19296

Gözat

Konu "Bilişim mühendisliği" ile LEE- Hesaplamalı Bilim ve Mühendislik-Yüksek Lisans'a göz atma
  • Öge
    Augmented superpixel based anomaly detection in hyperspectral imagery
    (Graduate School, 2024-07-01) Gökdemir, Ezgi ; Tuna, Süha ; 702211005 ; Computational Science and Engineering
    The detection of anomalies in hyperspectral images depends on several factors. Here, the spatial proximity of anomalies and confusion in the background image can create a bottleneck at the point of anomaly detection. Hyperspectral images are tensor data, in which each pixel contains both spatial and spectral information. These complex data structures pose significant challenges for traditional anomaly detection methods, which often struggle to account for the intricate relationships between the different spectral bands. In this thesis, a method called "Augmented Superpixel (Hyperpixel) Based Anomaly Detection in Hyperspectral Imagery" is proposed. This method aims to enhance the anomaly detection by leveraging advanced dimensionality reduction and segmentation techniques. Our approach begins by reducing the three-dimensional HSI data using methods such as high-dimensional model representation and Principal Component Analysis. This step simplifies the data while preserving critical spectral and spatial information. By capturing the most significant components of the data, these techniques help eliminate noise and irrelevant details, thereby making the subsequent analysis more focused and effective. We then applied segmentation methods such as Simple Linear Iterative Clustering and Linear Spectral Clustering to divide the image into distinct regions known as superpixels. Each superpixel is augmented with its first-order neighbors to form hyperpixels, which provide a richer context for anomaly detection. The augmentation process ensures that the local context is considered, thereby enhancing the ability to detect subtle anomalies that may be missed when examining individual superpixels in isolation. This neighborhood information is crucial for accurately identifying the boundaries of anomalies and distinguishing them from normal variations in the data. Finally, we applied the Local Outlier Factor algorithm to these hyperpixels to identify the outlier points that signify anomalies. The capability of the Local Outlier Factor to evaluate local density deviations enables it to accurately identify anomalies, even in densely populated or intricate backgrounds. The combination of these techniques ensures comprehensive and precise analysis that can handle the diverse characteristics of hyperspectral datasets. The proposed algorithm was tested using various hyperspectral image datasets and demonstrated good performance in detecting anomalies. By integrating dimensionality reduction, segmentation, and anomaly detection techniques, this method effectively manages the complexity of the hyperspectral data. This comprehensive approach allows for accurate identification of anomalies, even in challenging conditions where anomalies are closely packed or the background is complex. Through rigorous experimentation, the algorithm demonstrated robustness and reliability, making it a promising tool for hyperspectral image analyses. Its versatility and high accuracy across different datasets underline its potential for broad application in fields such as remote sensing, environmental monitoring, and urban planning. The ability to adapt to various anomaly characteristics and dataset structures makes this method a valuable addition to the toolkit for hyperspectral image-analysis techniques.
  • Öge
    Visualization based analysis of gene networks using high dimensional model representation
    (Graduate School, 2024-07-01) Güler, Pınar ; Tuna, Süha ; 702211009 ; Computational Science and Engineering
    Genetic studies have revolutionized our understanding of the biological mechanisms underlying health and disease. By exploring the intricate details of the human genome, researchers can identify genetic variations that contribute to various phenotypic outcomes. One of the key advancements in this field is gene network analysis, which examines the complex interactions between genes and how they regulate cellular processes. This approach provides a comprehensive view of the biological systems and uncovers the pathways involved in disease mechanisms. Genome-Wide Association Studies (GWAS) play a pivotal role among the methodologies utilized in gene network analysis. GWAS involves scanning the genome for slight variations, known as single nucleotide polymorphisms (SNPs), that occur more frequently in individuals with a particular disease or trait than in those without. By identifying these associations, GWAS helps pinpoint genetic factors contributing to disease susceptibility and progression, paving the way for personalized medicine and targeted therapeutic strategies. By integrating various variant analysis techniques, researchers can develop a deeper understanding of the genetic architecture of diseases, leading to significant advancements in diagnostics, treatment, and prevention. Gene network and pathway analyses are essential components of genetic studies, offering insights into genes' complex interactions and functions within a biological systems. However, both face significant computational challenges, mainly when dealing with high-dimensional genomic data. Analyzing vast datasets containing gene expression profiles and genetic variations demands sophisticated computational methods capable of handling their scale and complexity. Conventional statistical methods frequently require assistance to become effective, demanding complex computational approaches like data visualization, network modeling, and machine learning algorithms. In addition, the complexity of biological networks and pathways makes analysis even more complicated, necessitating the use of powerful computational tools to interpret regulatory mechanisms and simulate complex biological processes correctly. Overcoming these challenges is crucial for gaining deeper insights into gene networks and pathways, thereby advancing our understanding of their roles in health and disease. In pathway analysis, scientists employ data collected from many sources, such as Genome-Wide Association Studies (GWAS), to identify target genes and connect them to known pathways using Kyoto Encyclopedia of Genes and Genomes (KEGG) databases. However, pathway analysis presents major computing challenges, especially when large, high-dimensional genomic datasets are involved. Researchers have developed innovative methods such as High Dimensional Model Representation (HDMR), Chaos Game Representation (CGR), and visual analysis of DNA sequences based on a variant logic construction method called VARCH to overcome these challenges. By mapping genetic sequences into visual representations, these innovative approaches can help identify potential genetic markers and better understand biological processes. These computational methods must be included in gene network and pathway investigations to fully understand the complex architecture of genetic interactions and how they affect health and diseases. In this thesis, we harnessed three sophisticated computational methodologies: Chaos Game Representation, visual analysis of DNA sequences based on variant logic construction called VARCH, and High Dimensional Model Representation, each offering unique contributions to the variant analysis, respectively CGR, a prevalent technique in bioinformatics, translates genetic sequences into visually interpretable diagrams, clarifying complex structures and patterns in the sequences. On the other hand, VARCH converts sequences into a feature space, successfully capturing each aspect of their complexity and uncertainty. These techniques are effective instruments in our search for potential genetic markers that might help us distinguish between the patient and control groups in our investigation. Furthermore, we utilized HDMR for dimension reduction, an essential technique for simplifying the complex structure in high-dimensional genomic data. By condensing data dimensions, HDMR facilitated more efficient and accurate classification, enabling us to uncover sensitive genetic relationships and patterns that might have remained hidden otherwise. Integrating these computational techniques provided robust solutions for analyzing genetic data from the mTOR pathway, enriching our comprehension of the genetic mechanisms supporting various phenotypic outcomes. In our study, we begin on a mission to deepen our comprehension of the intricate genetic patterns intertwined with diverse phenotypic outcomes. Focusing on genetic data sourced from the mTOR pathway, we leveraged state-of-the-art computational methodologies to unravel hidden insights. Our primary objective was to assess the efficacy of CGR, VARCH, and HDMR in gene network analyses. As we analyzed the data, the results were quite compelling. Both CGR and VARCH methods demonstrated notable accuracy in genetic classification, with VARCH exhibiting a significant edge over CGR in terms of accuracy and sensitivity metrics. This superiority was underscored by VARCH's ability to considerably minimize binary cross-entropy (BCE) loss values, demonstrating the ability to reduce errors in predictions. However, we examined the computing overheads associated with each methodology in detail, providing insight into the challenging trade-off between computational complexity and accuracy. Despite the more significant parameters, VARCH's computational requirements were apparent, although its performance was better than CGR's. Our study demonstrates the potential of computational tools for unraveling gene complexities while also acting as an essential reminder of how crucial it is to overcome the complex environment of computational constraints carefully, helping researchers search for the best possible method selection and optimization.