Domain Adaptation Evaluation for Deep Image Segmentation

Author: ORCID icon
Lin, Kevin, School of Data Science, University of Virginia
Brown, Don, DS-Research, University of Virginia
Baek, Stephen, DS-Faculty Affairs, University of Virginia
Syed, Sana, MD-PEDT Gastroenterology, University of Virginia
Doryab, Afsaneh, EN-SIE, University of Virginia
Moskaluk, Christopher, MD-PATH Anatomical Pathology, University of Virginia

Advances in deep learning approaches for medical image segmentation show increasingly impressive results for disease detection. More specifically, these approaches have demonstrated a strong capability to detect and quantify cellular features despite significant differences in diseases, location, histopathology staining, size of the data, and imaging techniques. The goals of this dissertation are to develop a framework for medical image segmentation approaches, compare an existing domain adaptation approach with other methods, and create a new approach for future domain adaptation research. The first part of this dissertation focuses on evaluating the dice scores of the leading medical image segmentation model, U-Net, and using Monte Carlo Dropout to produce an entropy quantification metric to quantify and visualize areas where this model has difficulty in detection. Applicability of this approach is first proven with a baseline Brain Tumor Segmentation (BraTS) challenge 2021 dataset and then verified through segmentation evaluation on the private Eosinophilic Esophagitis (EoE) dataset. The created Monte Carlo Dropout U-Net maintains comparable dice scores on the EoE dataset and allows for a visualization of the entropy which highlighted detected cells and areas of interest. The second part of this dissertation focuses on extracting the entropy metric from the first part of this dissertation to separate observations into domains representing varying levels of entropy, using these domains to create a Multi-Domain Adversarial Network (MDAN), and comparing this MDAN performance to that of a Denoising Diffusion Probabilistic Models (DDPM). The motivation for the MDAN approach stems from the fact that adversarial network approaches are robust to the lack of available training data common in deep medical image segmentation. The third part of this dissertation introduces a new domain adaptation approach named Extremity-Ranked Domain Selection (ERDS) which ranks observations by their extremity and performs a full factorial experimental design to evaluate the impact of this domain choice on a MDAN dice score. An observation has high extremity if removing that observation's data in training has a large impact on the performance of the model. Observation extremity represents a new but important parameter in domain adaptation approaches for medical image segmentation. By successfully evaluating and creating domain adaptation techniques, both the medical and data science field benefit through greater understanding of how current deep medical image segmentation approaches detect diseases. Through these findings, future researchers can trust and leverage domain adaptation techniques on image segmentation applications.

PHD (Doctor of Philosophy)
Deep Learning, Medical Image Segmentation, Domain Adaptation, Bayesian Uncertainty, Experimental Design
Sponsoring Agency:
National Reconnaissance Office
All rights reserved (no additional license for public reuse)
Issued Date: