Deep understanding and generation of medical text and beyond
- Yuhao Zhang
- [Stanford, California] : [Stanford University], 2021
- Copyright notice
- Physical description
- 1 online resource
Also available at
- Zhang, Yuhao, author.
- Manning, Christopher D., degree supervisor.
- Altman, Russ, degree committee member.
- Potts, Christopher, 1977- degree committee member.
- Stanford University. Department of Biomedical Informatics.
- Human language text plays a pivotal role in medicine. We use text to represent and store our biomedical knowledge, to communicate clinical findings, and to document various forms of medical data as well as healthcare outcomes. While deep language understanding techniques based on neural representation learning have fundamentally advanced our ability to process human language, can we leverage this advancement to transform our ability to understand, generate and utilize medical text? If so, how can we achieve this goal? This dissertation aims to provide answers to these questions from three distinct perspectives. We first focus on a common form of medical text, biomedical scientific text, and study the long-standing challenge of extracting structured relational knowledge from this text. To handle the long textual context where biomedical relations are commonly found, we introduce a novel linguistically-motivated neural architecture that learns to represent a relation by exploiting the syntactic structure of a sentence. We show that this model not only demonstrates robust performance for biomedical relation extraction, but also achieves a new state of the art on relation extraction over general-domain text. In the second part of this work, we focus on a different form of medical text, clinical report text, and more specifically, the radiology report text commonly used to describe medical imaging studies. We study the challenging problem of compressing long, detailed radiology reports into more succinct summary text. We demonstrate how a neural sequence-to-sequence model that is tailored to the structure of radiology reports can learn to generate fluent summaries with substantial clinical validity. We further present a reinforcement learning-based method that optimizes this system for correctness, a crucial metric in medicine. Our system has the potential of saving doctors from repetitive labor and improving clinical communications. Finally, we connect text and image modalities in medicine, by addressing the challenge of transferring the knowledge that we learn from text understanding to understanding medical images. We present a novel method for improving medical image understanding by jointly modeling text and images in an unsupervised, contrastive manner. By leveraging the knowledge encoded in text, our method reduces the amount of labeled data needed for medical image understanding by an order of magnitude. Altogether, our studies demonstrate the great potential that deep language understanding and generation has in transforming medicine
- Publication date
- Copyright date
- Submitted to the Department of Biomedical Informatics
- Thesis Ph.D. Stanford University 2021