Abstract

In the biomedical field, downstream cell-level analysis is a crucial step for comprehensively characterizing cells' location, function, morphology, and phenotype, relying on the accurate identification of individual cells. Consequently, there is increasing interest in the high quality of identified cell boundaries for biologists or researchers. Currently, cell segmentation methods are categorized into two types: Image-based and RNA-based cell segmentation. However, these methods face significant challenges. For example, most image-based methods tend to identify cell nuclear boundaries rather than entire cell boundaries when only nuclear staining images are available, and their performance is affected by noise in images and training labels. RNA-based methods struggle to accurately delineate cell boundaries that align with actual images and usually require prior information like cell sizes and types. Additionally, while deep learning has improved segmentation accuracy, leading to the development of varied network architectures for performance optimization, these methods generally face challenges in stability, robustness, and generalization, when dealing with image modalities, tissue types, or cell types not previously encountered during training. This dissertation addresses these challenges by developing and evaluating efficient computational methods for cell segmentation, leveraging multi-modal and diverse data. This approach contributes significantly to the advancement of efficient and accurate methods in the field of cell segmentation. Specifically, the first topic is the development of GeneSegNet which integrates RNA spatial information and imaging information within a unified deep neural network to capture both visually plausible and biologically reasonable cell boundaries and is less susceptible to impact from a single information source on cell segmentation. The second topic is to conduct a systematic evaluation of 18 computational methods for cell segmentation. This evaluation emphasizes the overall segmentation effectiveness of cutting-edge deep learning methods, with a focus on segmentation benchmarks, scalability, and usability. Conducted across diverse tissue and cell types, as well as various imaging modalities, it aims to provide guidelines for the application and advancement of cell segmentation methods in real-world scenarios.

Library of Congress Subject Headings

Cytometry--Technique; Cytometry--Technological innovations; Deep learning (Machine learning)

Publication Date

1-5-2024

Document Type

Dissertation

Student Type

Graduate

Degree Name

Engineering (Ph.D.)

Department, Program, or Center

Engineering

College

Kate Gleason College of Engineering

Advisor

Yunbo Zhang

Advisor/Committee Member

Zhiqiang Tao

Advisor/Committee Member

Andres Kwasinski

Campus

RIT – Main Campus

Plan Codes

ENGR-PHD

Download

Share

COinS