Dissertation
Modern computing methods: harnessing AI, graph theory, and quantum computing
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2023
DOI: 10.25820/etd.007014
Abstract
The rapid increase in the amount of digital information worldwide comes with demands for various computational tasks. This presents a distinct opportunity to develop and apply computing techniques capable of revealing meaningful insights concealed within the intricate layers of information complexity. In order to effectively employ contemporary computing techniques in domain-specific computational tasks, it is essential to address the critical task of analyzing the requirements of the task and subsequently devising a suitable approach to tackle it. The emergence of machine learning, deep learning, graph theory, and lately quantum computing has been instrumental in addressing these challenges across various domains.
In this thesis, we first explore the practicality of using black-box machine learning models to provide useful insights into predictors and characteristics of hospital admissions for fall injuries in older patients. Local agnostic methods such as feature importance based on a mean decrease in the Gini index and logistic regression coefficients are crucial in the understanding of model-based learned predictions. On the other hand, there exist global explainable techniques that serve as universal probing tools. Specifically, model surrogate, drop-one column feature importance, and model ensemble, are capable of offering a comprehensive understanding across all machine learning models. By utilizing those tools, we successfully created a set of 18 distinct patterns, or signatures, that aid in identifying older individuals who are prone to experiencing fall-related injuries and may require hospital admission. In the second study, machine learning techniques were used to explore the factors that contribute to suicide deaths involving firearms. By combining the predictions of the top-performing models and producing an equivalent surrogate decision tree, combinations of high-risk factors predictive of suicide deaths in males by firearms were identified. This information can be used to develop an early detection tool and tailored prevention strategies for vulnerable individuals, particularly those residing in regions with high firearm ownership rates.
One of the major tasks in medical imaging is the segmentation of anatomical structures of interest. We developed graph-support deep learning tools to assist the segmentation of unruptured intracranial aneurysms. The first iteration of the tool allows for semi-automated segmentation with user inputs such as control points defining key locations of aneurysms and their vasculature. The second version of the tool called DeepLOGISMOS applies the self-training strategy to alleviate the lack of quality training labels and LOGISMOS post-processing module to improve segmentation masks. The proposed methods achieve high segmentation performance on five-fold cross-validation and superior detection sensitivity of 96% on an in-house dataset.
Last but not least, we explore the use of quantum computing, an emerging technology, in its feasibility to solve LOGISMOS single surface detection problem. Traditionally, this task requires solving for a minimum closed set in a directed graph. We proposed to transform the problem into a QUBO formulation and solve it using QAOA. We demonstrated that the proposed method can achieve comparable results to the classical approach on small simulation data. The results encourage further exploration of quantum computing in image processing and other demanding computing tasks.
Details
- Title: Subtitle
- Modern computing methods: harnessing AI, graph theory, and quantum computing
- Creators
- Nam Hoàng Lê
- Contributors
- Milan Sonka (Advisor)Fatima Toor (Committee Member)Colette Galet (Committee Member)Edgar Samaniego (Committee Member)Punam Saha (Committee Member)Xiaodong Wu (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Electrical and Computer Engineering
- Date degree season
- Autumn 2023
- DOI
- 10.25820/etd.007014
- Publisher
- University of Iowa
- Number of pages
- xx, 143 pages
- Copyright
- Copyright 2023 Nam Hoang Le
- Grant note
- This work was supported, in part, by the National Institutes of Health grant R01-EB004640, Society of Vascular and Interventional Neurology pilot grants, and the Bee Foundation. (82)
- Comment
This thesis has been optimized for improved web viewing. If you require the original version, contact the University Archives at the University of Iowa: https://www.lib.uiowa.edu/sc/contact/.
- Language
- English
- Date submitted
- 08/26/2023
- Description illustrations
- Illustrations, tables, graphs, charts
- Description bibliographic
- Includes bibliographical references (pages 136-143).
- Public Abstract (ETD)
Nowadays, there has been an increasing amount of data available, such as medical records and medical images, which can hold valuable insights if they can be uncovered. Finding appropriate steps to analyze and make sense of all this data, which can be complex, is important.
One approach to this challenge is by using computer techniques that learn from patterns called Machine Learning. In this thesis, machine learning models were used to predict if some elderly people are more likely than others to experience a hospital admission for a fall injury. Patterns in age, sex, comorbidities, and hospital information were analyzed in different ways to provide a ranking of the most impactful factors and how having multiple factors can affect potential outcomes.
In another study, we applied the same computing techniques to examine combinations of life, work, and relationship circumstances that are connected to future suicide deaths by firearms. By combining the predictions of multiple machine learning models, we identified groups of factors that are linked to this type of tragedy in the male population. This information could be used to create tools that help identify people who might be at risk and develop ways to prevent such incidents.
When it comes to medical images, there is a need for automatic tools for the identification of anatomical structures. We develop two approaches to tackle the problem of detecting unruptured brain aneurysms that potentially eliminate the need for manual work. The first method relies on LOGISMOS, a graph-based framework for finding surfaces in 3-D images, that provides an interactive platform allowing for the identification of aneurysm sacs and vessels with few user clicks. The second proposed method utilizes deep learning and the availability of public datasets to construct an automatic segmentation system. The two methods were tested on our in-house datasets provided by the University of Iowa Hospital and Clinics which yielded promising results, indicating the potential for practical application.
Lastly, we explored an emerging technology called quantum computing. Think of it as a unique way of processing information in the world of quantum mechanics. We developed an implementation of LOGISMOS that is solvable in quantum computers. In our simulation studies, the results showed that the quantum version of LOGISMOS is capable of delivering correct solutions for simple segmentation tasks. This is a promising step toward the development of quantum computing methods for medical image analysis.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9984546542702771
Metrics
4 File views/ downloads
54 Record Views