Thesis Title: Uncertainty Mitigation in Image-based Machine Learning Models for Precision Medicine 

Advisors: 

Dr. Jing Li, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Tech 

Thesis Committee: 

Dr. Jianjun Shi, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Tech 

Dr. Kamran Paynabar, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Tech 

Dr. Yajun Mei, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Tech 

Dr. Nathan Gaw, Department of Operational Sciences, Air Force Institute of Technology 

Date and Time: Thursday, July 21st, 2022, at 9:30 am (EDT) 

Location: Main 126 or meeting link below 

Meeting Link: https://gatech.zoom.us/j/98949968024?pwd=S21Bd0lhWjFmMUdjNFMvZzBpN2QzQT09 

Meeting ID: 989 4996 8024 

Passcode: 483282 

Abstract:  

Machine learning (ML) algorithms have been developed to build predictive models in medicine and healthcare. In most cases, the performance of ML models/algorithms is measured by predictive accuracy or accuracy-related measures only. In medicine, the model results are intended to guide physicians to make critical decisions regarding patient care. This means that quantifying and mitigating the uncertainty of the output is also very important as it will allow decision makers to know how much they can rely on the model output. 

My dissertation focuses on studying model uncertainty of image-based ML in the context of precision medicine of brain cancer. Specifically, I focus on developing ML models to predict intra-tumor heterogeneity of genomic and molecular markers based on multi-contrast magnetic resonance imaging (MRI) data for glioblastoma (GBM) – the most aggressive type of brain cancer. Intra-tumor heterogeneity has been found to be a leading cause of treatment failure of GBM. Devising a non-invasive approach to map out the molecular/genomic distribution using MRI helps develop treatment with high precision. My dissertation research addresses the model uncertainties due to high-dimensional and noisy features, sparsity of labeled data, and utility of domain knowledge.  

In the first study, we developed a Semi-supervised Gaussian Process with Uncertainty-minimizing Feature-selection (SGP-UF), which can incorporate selected unlabeled samples (i.e. unbiopsied regions of a tumor) in the model training, and integrate feature selection with a new criterion of seeking features that minimize the prediction uncertainty.  

In the second study, we developed a Knowledge-infused Global-Local data fusion (KGL) framework, which optimally fuses three sources of data/information including biopsy samples (labeled data, local/sparse), images (unlabeled data, global), and knowledge-driven mechanistic models.   

In the third study, we developed a Weakly Supervised Ordinal Support Vector Machine (WSO-SVM), which aims to leverage a combination of data sources including biopsy/labeled samples and unlabeled samples from the tumor and image data from the normal brain, as well as their intrinsic ordinal relationship.  

We demonstrate that these novel methods significantly reduce prediction uncertainty while at the same time achieving higher accuracy in precision medicine, which can inform personalized targeted treatment decisions that potentially improve clinical outcome.