Author

Ruoyu Li

ORCID Identifier(s)

0000-0002-6731-4261

Graduation Semester and Year

2020

Language

English

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Science

Department

Computer Science and Engineering

First Advisor

Junzhou Huang

Abstract

As the rise of graph neural networks, many deep learning frameworks have been extended to graph-structured data. The research in many diverse regimes have been tremendously reshaped, especially in areas like medical image understanding. When input data reach the scale of whole slides images (WSIs), the modeling becomes more challenging and we have to balance the trade-off between performance and efficiency. Furthermore, the theory of existing graph convolution has its own constraints which prevent learning robust graph representation on data that has diverse topological structure and are infeasible for graph sampling or coarsening. To tackle the problems we introduced a series of novel graph neural networks and techniques. For example, Adaptive Graph Convolutional Network (AGCN) combined the graph representation learning with graph learning and empower the network to learn features from hidden substructures on graph. Attentional-AGCN provided a node-to-graph attention scheme which does not only improve graph representation trained over large-scale images like WSIs, but also facilitate an intuitive explanation of the effectiveness of Attentional-AGCN. Besides, we introduced an end-to-end survival prediction framework based on WSI input directly, which delivered significant improvement on accuracy of survival analysis. Lastly, to mitigate the randomness introduced by patch sampling, a fast region-of-interest (RoI) search and detection approach is also introduced to images at WSI scale. MROID combined classification with detection into a unified framework to quickly narrow down candidates RoI proposals and gave a coarse-to-fine boundary refinement for generated RoIs. As summary, in this thesis, we reviewed and introduced our contributions to graph neural networks and the applications of our methods on large-scale biomedical data under- standing tasks.

Keywords

Graph neural network, Attention network, Region of interest detection, Whole slides image, Superpixel clustering, Drug discovery

Disciplines

Computer Sciences | Physical Sciences and Mathematics

Comments

Degree granted by The University of Texas at Arlington

29630-2.zip (14728 kB)

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.