You are cordially invited to my thesis defense on May 1st, 2025 (Thursday) at 1 PM.

Title: Data-efficient Machine Learning Methods in Healthcare Applications

Date: May 1st, 2025

Time: 1:00 PM – 3:00 PM

Location: Groseclose 303

Meeting Link: https://gatech.zoom.us/j/99117619626?pwd=A2i3RDe4F5q8sGKma4NIkze167YP7V.1

Zhiyang Zheng

Industrial Engineering PhD Candidate

H. Milton Stewart School of Industrial and Systems Engineering

Georgia Institute of Technology

Committee:

Dr. Jing Li (Advisor), H. Milton Stewart School of Industrial and Systems Engineering

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

Dr. Gian-Gabriel Garcia, H. Milton Stewart School of Industrial and Systems Engineering

Dr. Xiaochen Xian, H. Milton Stewart School of Industrial and Systems Engineering

Dr. Yi Su, Banner Alzheimer's Institute

Abstract:

Machine learning (ML) has emerged as a powerful tool in the healthcare sector, offering the potential to revolutionize diagnosis, treatment, and patient care. By analyzing complex datasets, ML algorithms can identify patterns and make predictions that can assist healthcare professionals in making informed decisions. However, the application of ML in healthcare presents unique challenges related to data efficiency. Key issues include the limited availability of accurately labeled medical data due to the cost and expertise required for labeling, the collection of more imaging modalities than necessary for some patients, increasing costs, and difficulties in sharing complex, multi-modality data across organizations while ensuring data security and patient privacy. To address these challenges, this thesis proposes three data-efficient ML methods tailored for specific healthcare applications. 

In Chapter 2, we propose integrating oral-anatomical knowledge with a Dense U-Net using regularized optimization for Cone Beam CT (CBCT) image segmentation and lesion detection, tackling the limited labeled data problem through a novel semi-supervised approach. In Chapter 3, we introduce an Uncertainty-driven Modality Selection (UMoS) framework to address inefficient data collection in predicting Alzheimer's disease (AD) progression from Mild Cognitive Impairment (MCI). This method sequentially adds data modalities based on model uncertainty, potentially reducing the need for costly imaging like PET scans while maintaining accuracy. In Chapter 4, we introduce Federated Variational Inference for Cross-Modal Learning (FedVICML) to address the challenge of incomplete multi-modality imaging data in federated settings for Alzheimer's Disease prediction. This framework leverages variational inference within federated learning to enable collaborative model training across sites without sharing raw data, improving predictive accuracy by probabilistically fusing available information and optimizing models locally.

Best wishes,