Item Details

abstract

Histopathological analyses play a central role in the characterization of biological tissues. Increasingly, technologies to create whole-slide imaging (WSI) of tissues, along with fast networks for data transfer and inexpensive storage, have made it possible to curate large databases of digitized tissue sections [1]. Furthermore, rapid advances in deep learning (a category of machine learning models) methods have enabled scientists to develop computer-assisted histopathological analysis methods on whole-slide images, ranging from basic applications such as nuclei detection [2] and mitosis detection [3] to more advanced applications such as tumor grading [4]. By and large, the success of these developments provides substantial interest in deep learning as an essential tool for histopathological analyses for clinical and biomedical applications.Despite successful application to various diagnostic and prognostic problems [1, 5], computational pathology methods still mostly rely on painstakingly annotated nuclei, cells, and tissue structures [6-10]. This is driven primarily by the prevalence of supervised methods in more generalized computer vision applications. Unlike general computer vision applications, the reliance on annotations heavily limits research in computational pathology, as annotations must be performed by expert pathologists. However, annotations are labor-intensive and often subject to significant inter- and intra- reader variability. The recently published vision transformer [11], which has been heavily cited, relies on a 303 million image dataset with annotations for each image. Creating such a dataset is impractical for medical applications, as medical image datasets are vanishingly small compared to general-purpose datasets. Consequently, recent high-profile publications in computational pathology have started moving away from fully-supervised methods to semi- and weakly-supervised methods [12, 13]. The preliminary work of this proposal coincides with and arguably pioneers this semi- and weakly- supervised trend. Furthermore, this proposal uniquely adapts semi- and weakly- supervised methods in novel manners providing frameworks for knowledge discovery and validation of existing practices in clinical and diagnostic pathology. The former, knowledge discovery, explicitly refers to novel, automatically learned, interpretable imaging biomarkers for disease despite having no tissue-level annotations. Finally, it culminates with a novel approach to computational pathology as a whole, in which meaningful features can be learned from WSIs without needing any annotations. We present these semi- and weakly- supervised methods in two clinical applications. Chapters 1 and 2 focus on minimizing tissue-level annotations via application to colorectal cancer (CRC). Chapters 3, 4, and 5 abandon tissue-level annotations in favor of whole slide labels (i.e., a diagnosis or some continuous variable assigned to a biopsy) and applies a novel framework to tuberculosis (TB). Finally, Chapter 6 abandons annotations and labels completely in favor of an unsupervised model for learning meaningful, compact representations of WSIs for posterior classification and regression tasks.

subject

cancer

computational pathology

deep learning

imaging biomarkers

tuberculosis

weak supervision

contributor

Tavolara, Thomas Erol (author)

Gurcan, Metin N (committee chair)

Topaloglu, Umit (committee member)

Weaver, Ashley (committee member)

Beamer, Gillian (committee member)

Hsi, Eric (committee member)

date

2022-09-17T08:35:43Z (accessioned)

2023-09-16T08:30:07Z (available)

2022 (issued)

degree

Biomedical Engineering (discipline)

embargo

2023-09-16 (terms)

identifier

http://hdl.handle.net/10339/101253 (uri)

language

en (iso)

publisher

Wake Forest University

title

Deep Learning for Computational Pathology

type

Dissertation