Graduation Year

2022

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Computer Science and Engineering

Major Professor

Lawrence O. Hall, Ph.D.

Co-Major Professor

Dmitry B. Goldgof, Ph.D.

Committee Member

Shaun Canavan, Ph.D.

Committee Member

Ashwin Parthasarathy, Ph.D.

Committee Member

Robert Gatenby, M.D.

Keywords

CNN, Ensemble learning, Medical image analysis, Generalization, Shortcut learning

Abstract

Artificial Intelligence (AI) is “The science and engineering of making intelligent machines, especially intelligent computer programs”. Artificial Intelligence has been applied in a wide range of fields including automobiles, space, robotics, and healthcare.

According to recent reports, AI will have a huge impact on increasing the world economy by 2030 and it's expected that the greatest impact will be in the field of healthcare. The global market size of AI in healthcare was estimated at USD 10.4 billion in 2021 and is expected to grow at a high rate from 2022 to 2030 (CAGR of 38.4%). Applications of AI in healthcare include robot-assisted surgery, disease detection, health monitoring, and automatic medical image analysis. Healthcare organizations are becoming increasingly interested in how artificial intelligence can support better patient care while reducing costs and improving efficiencies.

Deep learning is a subset of AI that is becoming transformative for healthcare. Deep learning offers fast and accurate data analysis. Deep learning is based on the concept of artificial neural networks to solve complex problems.

In this dissertation, we propose deep learning-based solutions to the problems of limited medical imaging in two clinical contexts: brain tumor prognosis and COVID-19 diagnosis. For brain tumor prognosis, we suggest novel systems for overall survival prediction of Glioblastoma patients from small magnetic resonance imaging (MRI) datasets based on ensembles of convolutional neural networks (CNNs). For COVID-19 diagnosis, we reveal one critical problem with CNN-based approaches for predicting COVID-19 from chest X-ray (CXR) imaging: shortcut learning. Then, we experimentally suggest methods to mitigate this problem to build fair, reliable, robust, and transparent deep learning-based clinical decision support systems. We discovered this problem with CNNs and using Chest X-ray imaging. However, the issue and solutions generally apply to other imaging modalities and recognition problems.

Share

COinS