Leveraging Machine Learning for Improved Data Interpretation and Decision Making Under Uncertainty

Abstract

Machine learning methodologies can provide viable solutions to a wide range of complex real-world problems. However, in practice, issues such as data noise, class imbalance, and uncertainty inherent in datasets and models are frequently encountered. This dissertation investigates machine learning strategies for crowdsourced labeling, medical diagnosis, driver distraction detection, and biological image classification. It proposes a soft-weighted aggregation algorithm for crowdsourced labels that uses annotators' consistency in relation to a reference classifier to determine their reliability, demonstrating a 20% improvement in accuracy over ten existing label aggregation techniques across ten distinct datasets when only three workers are present. The dissertation also describes a hierarchical classification methodology that uses disease taxonomy to improve chest radiograph classifications, demonstrating a 15% improvement in accuracy and AUC score across different diseases when the proposed approaches are used. It presents a cascaded multi-planar convolutional neural network design for improved thalamic nuclei segmentation in MRI scans. Transfer learning is used in two different applications: identifying driver distractions in images and classifying primary cilia in microscopy data. The efficacy of these methodologies is supported by experimental results on real-world datasets. The approaches and findings in this dissertation contribute to machine learning advancements and may provide practical solutions to some of the challenges encountered in various sectors, such as data science, radiology, and road safety.