Graph and Vision Based Learning for Medical Image Analysis: Applications in Brain Networks and Omics Phenotyping Jiong Chen

Format:

Book

Thesis/Dissertation

Author/Creator:

Chen, Jiong, author.

Contributor:

University of Pennsylvania. Bioengineering., degree granting institution.

Language:

English

Subjects (All):

0202.

0574.

0308.

Local Subjects:

0202.

0574.

0308.

Genre:

Academic theses

Physical Description:

1 electronic resource (221 pages)

Contained In:

Dissertations Abstracts International 88-03B

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2026

Language Note:

English

Summary:

Medical image analysis plays a critical role in advancing our understanding of the heterogeneity and complexity of human physiology and disease. With the rapid evolution of biomedical technologies, medical data are no longer confined to conventional anatomical imaging, but increasingly encompass high-dimensional, multi-modal, and multi-scale representations that challenge existing analytical frameworks. This brings attentions for computational frameworks that are both flexible across data modalities and interpretable across biological scales. In neuroimaging, functional magnetic resonance imaging (fMRI) produces continuous blood-oxygen-level-dependent (BOLD) signals for each voxel across time, enabling the characterization of functional connectivity patterns among distributed brain regions. Traditional analysis methods, however, often rely on static or population-averaged representations that prevent observation of individual variability and higher-order network structure. This dissertation develops graph-based and topological frameworks for fMRI analysis that model brain connectivity as structured networks, capturing both global and individual-specific properties as well as higher-order interactions. By integrating tools from topological data analysis and graph learning, the proposed methods disentangle shared and subject-specific connectivity components, providing interpretable insights into brain functional organization at both macroscopic and mesoscopic scales. Beyond neuroimaging, modern biomedical analysis increasingly involves data that are not images in the classical sense but can nonetheless be interpreted through a visual perspective. High-dimensional cytometry data, for example, are typically analyzed as tabular measurements through manual gating or clustering, leading to subjectivity and limited reproducibility. This dissertation reframes cytometry phenotyping as a computer vision-based problem, introducing segmentation-based approaches for automated, consistent, and interpretable cell population identification. These methods bridge human expert reasoning and algorithmic rigor, enabling scalable and reproducible immune phenotyping across large cohorts. At the cellular and tissue level, recent advances in spatial transcriptomics further expand the concept of medical imaging by providing spatially resolved gene expression profiles in situ. Such data offer great opportunities to study tissue architecture and cellular interactions, particularly when coupled with cellular imaging modalities such as histology, nuclear staining, or membrane staining. However, existing computational methods are often limited to single-view or modality-specific analyses, restricting their ability to generalize across datasets and biological contexts. This dissertation proposes vision-based and multi-modal learning frameworks for spatial transcriptomics analysis, particularly single-cell segmentation approaches that integrate transcriptomic and imaging modalities together. By leveraging shared representations across modalities, the proposed methods enhance scalability, robustness, and biological interpretability. Collectively, this dissertation introduces a unified set of graph- and vision-based computational frameworks for multi-scale medical image analysis, spanning brain networks, single-cell measurements, and spatial transcriptomics imaging. The contributions advance methodological foundations in representation learning and network modeling while addressing practical challenges in translational biomedical research. Through applications in neuroscience and cellular phenotyping, this work demonstrates how principled computational approaches can yield interpretable insights into complex biological systems and support more precise, reproducible, and scalable biomedical analysis

Notes:

Source: Dissertations Abstracts International, Volume: 88-03, Section: B.

Advisors: Shen, Li; Kim, Dokyoon Committee members: Shou, Haochang; Huang, Zhi

Ph.D. University of Pennsylvania 2026

Includes bibliographical references

Local Notes:

School code: 0175

ISBN:

9798189294204

Access Restriction:

Restricted for use by site license