Deep Generative Models for Medical Images and Beyond
Open Access
- Author:
- Xue, Yuan
- Graduate Program:
- Information Sciences and Technology
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- April 09, 2021
- Committee Members:
- Sharon Xiaolei Huang, Dissertation Advisor/Co-Advisor
Sharon Xiaolei Huang, Committee Chair/Co-Chair
Zihan Zhou, Committee Member
Ting Wang, Committee Member
Robert Collins, Outside Member
James Duncan, Special Member
Mary Beth Rosson, Program Head/Chair - Keywords:
- Deep Learning
Computer Vision
Generative Models
Medical Image Analysis - Abstract:
- Deep generative models such as generative adversarial networks (GANs) or variational autoencoders (VAEs) have been extensively studied in unsupervised image generation tasks where given training data, the goal is to try and generate new samples from the same distribution. Despite such research efforts, components of such generative models, especially GANs, have rarely been integrated into supervised learning scenarios such as in classification, segmentation, and regression tasks in the literature. To improve the generality and applicability of deep generative models, in this dissertation, we show the potential of integrating components of generative models into a variety of supervised learning tasks for improved performance. We propose several advanced deep learning based generative methods that are complementary to traditional supervised learning methods, for different medical image analysis applications as well as architecture design applications which achieve state-of-the-art performances in our experiments. First, we present an image segmentation method that consists of a segmentor (i.e., generator) network and a critic (i.e., discriminator) network, which is trained in an adversarial learning fashion, so that feedback from the critic network can help the segmentor generate accurate and realistic segmentations. We also discuss other segmentation tasks, including 3D organ segmentation and infant video segmentation. Then, we present a multimodal recurrent model with attention for the automatic generation of medical reports given X-Rays images. We also propose several deep generative model architectures with structural integrity for indoor wireframe scene rendering and automated floor plan generation. Finally, we propose a series of synthetic augmentation models that generate synthetic images and then selectively choose high-quality synthetic images to augment training sets and improve histopathology image classification results.