Boston University
Context-aware digital pathology: enhancing whole slide image analysis through global-local feature learning
Abstract
dc:description.abstractComputational pathology, which entails the analysis of digitized pathology slides, is gaining increased attention over the past few years. The sheer size of a single whole slide image (WSI) typically can exceed a gigabyte, so traditional image analysis routines may not be able to fully process all this data in an efficient fashion. Modern machine learning methods such as deep learning have allowed us to make great progress in terms of analyzing WSIs including disease classification, tissue segmentation, mutation prediction, and spatial profiling of immune infiltration. In this context, weakly-supervised deep learning (WSDL) gained increasing interest because it utilizes slide- or patient-level labels that are routinely collected for clinical purposes. This approach has achieved state-of-the-art performance in various tasks within digital pathology analysis. Traditional methods typically rely on manually crafted, domain-specific features, whereas WSDL methods can learn representations without manually designed features. However, WSI is unique in various ways and presents its own set of computational challenges. In this thesis, we developed deep neural networks to learn efficient, context-aware representations of WSIs that can perform various computational pathology tasks, such as kidney fibrosis estimation, cancer diagnosis, and survival outcome analysis. First, we study the advantages of leveraging merits from Global-Local Feature Learning, which enables the seamless integration between global contextual information and necessary local fine details of the WSI, to ensure accurate WSI-level classification. We propose a novel Global-Local Pathology (GLPath) network. It combined the features learned at the global level of the WSI along with the ones learned from local high-resolution image patches from the WSI. GLPath can effectively and efficiently integrate global context and local high-resolution fine structures, yielding state-of-the-art on Interstitial fibrosis and tubular atrophy (IFTA) estimation on a renal biopsy. Second, based on insights from global-local feature learning in the IFTA estimation task, we propose redefining the WSI as a graph-based feature learning model. To demonstrate the effectiveness of this approach, we propose a novel context-aware, spatially-resolved patch-based graph-transformer pathology (GTP) network. This network hierarchically aggregates patch-level histology features to capture both local and global-level topological structures within the WSI microenvironment. The GTP network exhibits state-of-the-art performance in distinguishing between lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) from adjacent non-cancerous tissue in Non-Small Cell Lung Cancer (NSCLC) analysis. Lastly, we focus on the most challenging yet practical task of cancer survival prognostication. Building on our progress in representation learning of the entire WSI, we propose a multimodal feature learning to overcome data heterogeneity gap in incorporating biological priors such as genomic measurements. An intuitive and novel approach is proposed and studied in this thesis, i.e., Pathology Mixer (PathMixer) network. PathMixer is an attention-based fusion architecture that integrates a graph representation of pathology images with gene expression data and concomitantly learns from the fused information to predict patient-specific survival.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Zheng, Yi
- Advisor dc:contributor.advisor
-
- Betke, Margrit
- ORCID dc:identifier.orcid
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53130