Data-Driven General Purpose Foundation Models for
Computational Pathology
Name
lu-mingylu-phd-eecs-2025-thesis.pdf
Description
Thesis PDF
Size
92.88 MB
Format
Adobe PDF
Checksum (MD5)
e41eb6523bacdd1769ee86f897123607
Author(s)
Lu, Ming Yang (Max)
Advisor(s)
Mahmood, Faisal
Date Issued
February 2025
Publisher
Massachusetts Institute of Technology
Abstract
The field of computational pathology has undergone a remarkable transformation in recent years. Researchers have leveraged supervised and weakly-supervised deep learning with varying degrees of success to address a wide range of tasks, including cancer subtyping and grading, metastasis detection, survival and treatment response prediction, tumor site-of-origin identification, mutation prediction, biomarker screening, and more. However, traditional task-specific models often require extensive labeled data and struggle to generalize across diverse pathology tasks. This limitation motivates the exploration of foundation models, which promise a more scalable, versatile solution by learning broad representations that can be adapted to various downstream applications. In this thesis, I will investigate the capabilities and limitations of data-driven foundation models in computational pathology. Specifically, I will explore two frameworks for developing general-purpose encoder models for pathology images: one using paired image-text data, and another leveraging self-supervised learning on large-scale unlabeled images. Additionally, I will examine downstream applications of these foundation models, including zero-shot transfer to gigapixel whole slide images and the development of an interactive multimodal AI assistant for pathologists.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
