LEE- Bilgisayar Mühendisliği-Doktora

Bu koleksiyon için kalıcı URI

http://hdl.handle.net/11527/19208

Gözat

Yazar "Bilgen, İsmail" ile LEE- Bilgisayar Mühendisliği-Doktora'a göz atma
  • Öge
    Machine-learning approaches for neurological disorder diagnosis from genomic and neuroimaging data
    (Graduate School, 2024-04-05) Bilgen, İsmail ; Töreyin, Behçet Uğur ; 504132516 ; Computer Engineering
    Millions of individuals across the globe are impacted by neurological disorders. These disorders contribute to 6.3% of the global burden of disease. In addition, research indicates that as of 2005, approximately 25 billion people worldwide, irrespective of age, gender, education level, or nationality, were affected by neurological disorders. These conditions are associated with an estimated annual toll of 6.8 million fatalities. Neurological disorders encompass a wide range of diseases affecting the nervous system exhibiting diverse etiologies, manifestations, and outcomes. The diagnosis of brain disorders involves a comprehensive assessment of physical and neurological examinations, neuropsychological tests, detailed medical histories, neuroimaging outputs, and in some cases, specific blood markers and genetic testing. In particular, neuroimaging techniques such as functional and structural MRI, CT, and PET, focusing on visualizing and mapping the brain's structure and activity, are crucial for diagnostics. Additionally, genetic testing may also be considered for certain neurological disorders with a known genetic component. Despite advancements in measuring methods and the comprehensive assessment of a variety of information, traditional diagnostic methods often fall short of providing timely and accurate assessments. The complex structure of the brain and the yet-to-be-fully elucidated genetic mechanisms underlying brain development complicate the development of effective diagnostic and treatment methods for neurological disorders. This gap encourages research groups to develop diagnostic techniques and discover new biomarkers by leveraging genetic and neuroimaging data. To address the identified research gap, this thesis primarily tackles two key research objectives: unraveling the genetic foundations of neurological disorders and advancing diagnostic methods, by utilizing the machine learning. Machine learning, a subfield of artificial intelligence, enables computer systems to learn from data and perform complex tasks by enhancing their ability to learn from data. Machine learning methods have been successfully applied in various fields. Especially, machine learning methods, equipped to analyze extensive datasets, have great potential in advancing and exploring new diagnostic methods for neurological disorders. These benefits make ML a substantial tool for various tasks including identification of novel biomarkers, unveiling associations between genetic factors and neurological conditions, and improving diagnostic skills. In addition to a deeper understanding of the underlying mechanisms of neurological disorders, harnessing the power of machine learning offers manifold benefits such as early detection of neurological disorders, enhanced diagnosis accuracy, and the development of more effective treatments. Neurological disorders can result from plenty of various factors including genetic abnormalities, developmental issues, infections, immune-mediated responses, structural abnormalities, metabolic disturbances, or a combination of these factors are commonly known causes of brain disorders. However, at the heart of neurological disorders lies a genetic landscape that guides the complex choreography of brain development. The dysregulation of genes can result in the modification of circuitry and communication patterns within the brain, contributing to the diverse array of symptoms seen in neurodevelopmental disorders. Therefore, the study conducted in Chapter 3 focuses on inferring the gene regulatory networks in humans to provide useful insights into unraveling the underlying reasons for neurological disorders. To tackle this challenge, we introduce Autoencoder for Gene Regulatory Network (AEGRN), a novel method utilizing Autoencoder (AE), in the pursuit of inferring gene regulatory networks solely from gene expression data. AEGRN is a fully unsupervised approach. It employs Autoencoders to approximate the function that reconstructs gene expression inputs while adhering to defined constraints. Samples in the gene expression data comprise mRNA levels at a specific moment. Therefore, alterations in gene expression levels across different samples provide insights into regulatory interactions. The trained Autoencoder learns how to encode input data into a hidden layer and decode it back. Ultimately, the trained AE network is exploited to unveil regulatory interactions between gene pairs. Here, two approaches are proposed: examining trained AE network weights and analyzing the AE model output after partial interventions. In this study, we aim to explore both direct and indirect regulatory interactions among genes since genes can be regulated as a result of a nonlinear combination of several regulatory steps. Consequently, the inferred regulatory interactions encompass both direct and indirect relationships, contributing to a comprehensive understanding of gene regulatory networks. Neurological disorders can impact the morphology of anatomical brain regions to varying extents. Recognizing the distinctive morphological features associated with a specific brain disorder can enhance diagnostic accuracy and provide insights into how neuroanatomical changes relate to function and cognition. The study conducted in Chapter 4 aims to apply different machine learning pipelines using morphological features, in the diagnosis of ASD as a neurological disorder. Autism spectrum disorder (ASD) impacts brain connectivity at various levels, posing a challenge for machine learning diagnostic models to non-invasively discern these effects through magnetic resonance imaging (MRI) due to the heterogeneity of ASD. Existing network neuroscience research has predominantly focused on functional (from functional MRI) and structural (from diffusion MRI) brain connectivity, potentially overlooking relational morphological changes between brain regions. Notably, machine learning studies are scarce for ASD diagnosis utilizing morphological brain networks derived from traditional T1-weighted MRI. To address this gap, the study conducted in Chapter 4 leverages crowdsourcing to apply various machine learning pipelines for neurological disorder diagnosis, specifically targeting ASD diagnosis through cortical morphological networks derived from T1-weighted MRI. Participant teams were provided a training dataset and requested to develop their machine learning pipelines to predict the ASD status of patients. The test set comprises public and private parts. The teams could only assess the performance of their pipeline on public test data. However, the final evaluation considered both public and private test data, using the metrics of accuracy, sensitivity, and specificity. Teams were ranked based on each metric individually, and the overall ranking was calculated by averaging rankings in all metrics. Consequently, 20 ML pipelines obtained from the top 20 teams are examined in terms of three categories: preprocessing techniques, feature selection/dimensionality reduction methods, and learning models. The top-ranked team achieved 70% accuracy, 72.5% sensitivity, and 67.5% specificity, while the second-ranked team achieved 63.8%, 62.5%, and 65%, respectively. The utilization of participants in a competitive machine learning environment facilitated the exploration and benchmarking of a broad spectrum of ML methods for ASD diagnosis using cortical morphological networks. Resting-state functional MRI (rs-fMRI) stands out as a key method, measuring intrinsic neural activity and providing potential biomarkers for neurological disorders through functional connectivity. A growing body of evidence suggests that neurological disorders induce changes in functional connectivity. However, traditional machine learning and deep learning frameworks face limitations due to their inability to effectively utilize the topological structure inherent in brain connectivity graphs. The study conducted in Chapter 5 tackles this challenge by employing a graph neural network (GNN) that leverages brain graphs derived from functional connectivity data. Moreover, this study addresses a crucial gap in prior research, in which the emphasis has been on feature selection and extraction to increase diagnosis accuracy and the significant impact of poor data quality on the efficiency of machine learning models has been overlooked. To fill this gap, we propose a fully unsupervised approach, Influence-Based Detection of Opposite Samples (IDOS), aimed at systematically estimating the influence of individual subjects within the functional brain network on the model's ability to diagnose Autism Spectrum Disorder (ASD). IDOS exploits gradient to calculate the influence of the training samples. Ultimately, it enables the detection of opposite samples within training data and enhances the diagnosis performance by excluding them. IDOS is applied in ASD diagnosis task from functional connectivity data using graph-based GNN models: GCN and DIFFPOOL. The significance of this research lies in its potential to transform the diagnostic landscape of neurological disorders. Current ASD diagnostic methods often rely on time-consuming and subjective evaluations, leading to delayed interventions and personalized treatment plans. Through the application of machine learning and data-driven techniques, this thesis aims to establish more objective and efficient diagnostic tools, facilitating timely interventions and enhancing outcomes for individuals with ASD. Furthermore, the research aims to bridge the existing gap between machine learning and neuroscience, contributing to a deeper understanding of the biological mechanisms underlying neurological disorders. This interdisciplinary approach holds promise for enriching both the fields of machine learning and neuroscience, paving the way for collaborative research in the future. This thesis marks a significant advancement in applying machine learning to propel the diagnosis of neurological disorders. Focused on ASD and employing various advanced machine learning techniques, the study aspires to make a meaningful impact on healthcare, ultimately improving the accuracy of diagnosing not only ASD but potentially other neurological disorders in the future.