Date of Award

2025

Document Type

Thesis (Ph.D.)

Department or Program

Quantitative Biomedical Sciences

First Advisor

Saeed Hassanpour

Abstract

Precision medicine is a data-driven approach that tailors treatments to individualized need. Emerging in the late twentieth century, it emphasizes customizing therapy based on each person’s unique characteristics and became a major national focus in the US starting in 2015.

Cancer remains a leading cause of death worldwide. With high mortality rates and resource- intensive treatment process, it burdens the healthcare systems. One of the inherent challenges in cancer care is its biological heterogeneity, making diagnosis and treatment planning highly complex and calls for personalized cancer management.

Today, diverse data sources are available to support cancer management and treatment decisions. Despite this, cancer management largely depend on manual interpretation of medical images and consideration of a limited number of risk factors. Several challenges remain: (a) it is infeasible for clinicians to comprehensively integrate all available data for decision-making; (b) patients respond differently to the same treatments, leading to uncertainty in outcomes and suboptimal prognoses; and (c) cancer diagnosis and care are resource intensive and heavily reliant on specialist expertise, which may not be readily accessible in resource-limited seƫngs.

This dissertation presents three novel pipelines that advance precision oncology through multimodal deep learning frameworks.

The first pipeline targets non-small cell lung cancer (NSCLC), developing a multimodal model that combines histology whole-slide images, clinical data, and next-generaƟon sequencing data to predict resistance to osimertinib. The model demonstrated strong performance in risk prediction, stratification, and interpretability, offering clinical value in decision-making.

The second pipeline introduces a contrasƟve learning-enhanced framework for survival prediction in lower-grade glioma patients. By integrating histopathology images, somatic mutation, and clinical variables, the model achieved a high c-index and robust risk stratification, outperforming models without contrastive learning.

The third pipeline addresses diagnostic and grading challenges in bone lesion classification using a contrastive learning-enhanced model trained with histology and X-ray images. It achieved strong performance using only X-ray images at test time, reducing reliance on invasive biopsy and histology imaging. Together, these studies demonstrate the value of contrastive and multimodal learning in improving clinical prediction, contributing novel model architectures, training strategies, and applications that expand the scope and impact of precision cancer care.

Download

Share

COinS