Access Type
Open Access Dissertation
Date of Award
January 2021
Degree Type
Dissertation
Degree Name
Ph.D.
Department
Computer Science
First Advisor
Suzan D. Arslanturk
Abstract
Deep learning has become an increasingly popular trend in recent years with applications in different domains including healthcare and medicine using numeric, spatially 2D and 3D, and time-series based video and audio data. However, data collection being the principal bottleneck for advancement in the life sciences, particularly in genomics, engineering, and healthcare, has led to incomplete and insufficient data, resulting in non-optimal deep learning model performances. Here, we propose solutions to overcome such data limitations when high dimensional medical data with limited sample sizes are present. These solutions include but not limited to multi-class learning, feature compression (\eg, auto-encoder), data augmentation, missing modality imputation, few-shot learning, and unsupervised learning to solve various problems in bioinformatics and computer vision domains.
Specifically, the chapters are summarized as follows: In Chapter 2, we have identified jointly important biomarkers across ovarian, prostate, and breast cancers by leveraging the biological and molecular similarities across hormonally driven cancers through an explainable multi-label classification auto-encoder. The proposed cross-cancer learning framework has enriched the study population and resulted in novel reproducible biomarkers for better diagnosis and treatment planning. The meta-analysis based on this chapter helps to distinguish mechanisms and pathways that are common across the three cancers. In Chapte 3, novel morphological biomarkers associated with cancer subtypes are explored for better diagnosis and prognosis through data augmentation and patch generation methods on histopathology images. In Chapter 4, a novel cancer progression prediction system is proposed by flexibly integrating collective information available through multiple studies with different cohorts and incomplete data types using innovative deep learning models and algorithms. Finally, in Chapter 5, a novel ``zero-shot'' upsampling method has been proposed on 3D point clouds to generate a denser representation of the input 3D shape to provide the operative visualization or bioprinting with higher quality and less time cost for the usage in medical or other domains.
Recommended Citation
Zhou, Kaiyue, "Novel Deep Learning Methods For Medical Applications With Limited Data" (2021). Wayne State University Dissertations. 3558.
https://digitalcommons.wayne.edu/oa_dissertations/3558