ORCID Identifier(s)


Graduation Semester and Year




Document Type


Degree Name

Doctor of Philosophy in Computer Science


Computer Science and Engineering

First Advisor

Junzhou Huang


Nuclei segmentation is a fundamental but challenging task in histopathology image analysis. For semantic segmentation of nuclei, Convolutional Neural Network (CNN), and Vision Transformer (VT) models give very promising results. However, to successfully train fully-supervised CNN and VT models we need significant amount of annotated data which is highly rare in biomedical domain. Also, collecting an unannotated histopathology dataset first, and then manually doing pixel-level labeling is expensive, time-consuming and tedious process. Therefore, we require to discover a way for training nuclei segmentation models with unlabeled datasets. In this thesis, I present my work towards solving this critical problem by utilizing Adversarial Learning, Self-Supervised Learning (SSL), and Diffusion Models. Thus, my approaches can be summarized into three directions: 1) employing adversarial learning based unsupervised and semi-supervised domain adaptation techniques to solve nuclei segmentation problem for unannotated datasets; 2) proposing SSL based approaches for pre-training VT models with unannotated image dataset; 3) introducing Denoising Diffusion Probabilistic Model (DDPM) based approach for pre-training nuclei segmentation model with large-scale histology image dataset. In the first approach, I apply Unsupervised Domain Adaptation (UDA) and Semi-Supervised Domain Adaptation (SSDA) with the help of another labeled dataset that may come from another organs or sources. Later, I extend the model by utilizing an adversarial learning incorporated reconstruction network to translate the source-domain images to the target domain for further training. Then, in my second approach, I introduce a novel region-level SSL based framework for pre-training semantic nuclei segmentation model with a large-scale unannotated histopathology image dataset extracted from Whole Slide Images (WSI). Additionally, I propose hierarchical, scale, and transformation equivariance loss to reduce the disagreements among predictions. Finally, in the third approach, I utilize DDPM for extracting discriminative and powerful features. Then, I combine a generation module, a discriminator, and scale loss with DDPM for effective label-efficient SSL based pre-training. Extensive and comprehensive experiments demonstrate the superiority of the proposed methods over the baseline models.


Nuclei segmentation, Adversarial learning, Self-supervised learning, Denoising model


Computer Sciences | Physical Sciences and Mathematics


Degree granted by The University of Texas at Arlington