Whole Slide Image Classification Using Deep Learning.
Loading...
Embargo End Date
2026-02-20
ICR Authors
Authors
Fourkioti, O
Document Type
Thesis or Dissertation
Date
2025-02-20
Date Accepted
Abstract
The digital transformation of histopathological assessment through Whole Slide Imaging (WSI) has
sparked a revolution in digital pathology, primarily propelled by advancements in artificial
intelligence (AI) and deep learning (DL). This thesis seeks to the contribute to this ongoing
technological evolution by developing, and evaluating methods that broaden the current capabilities
of pathological diagnostics. Specifically, we focus on WSI classification tasks, using solely the
visual information present in the histology images, without incorporating any other data
modalities.
However, WSI classification models struggle to generalise on new, unseen data due to the limited
availability of diverse and standardized histopathology datasets. In the first part of our thesis,
we introduce a new a diverse soft tissue sarcoma (STS) dataset. The STS dataset comprises 1163 WSIs
from 303 patients collected from the Royal Marsden Hospital (RMH). This dataset, spanning over 12
subtypes, is intended to be publicly accessible, fostering further studies in the field of WSI
analysis and contributing to the development of models with enhanced generalizability.
A recent trend emerging in WSI classification is adopting graph-based approaches, where every tile
within a WSI represents the central node of a graph and its spatially closest neighbouring tiles
constitute the surrounding nodes. Many WSI approaches commonly use a fixed number of neighbours for
constructing the graphs. In the second part of our thesis, we explore whether the contextual
information provided by neighbouring tiles enhances or hinders the network's performance in
identifying patterns and features relevant to the performance of the WSI classification models.
Through systematic variations in the content and arrangement of neighbouring tiles, we explore this
aspect using two histopathology datasets.
In the third part of our thesis, we present Context-aware Multiple Instance Learning. (CAMIL)
addresses the complex challenges of intra- and intertumoral heterogeneity by leveraging
dependencies among tiles within WSIs. Our findings, validated on two widely used histopathology
datasets, namely camelyon16 and TCGA- NLSC, demonstrate substantial enhancements in classification
performance.
In the fourth part our thesis, we introduce an instance-based loss inspired by semi-supervised
anomaly detection (AD). Leveraging the available slide-level labels, the AD-loss plays a crucial
role in discerning between normal and anomalous tiles in WSIs. When dealing with normal slides, the
AD-loss exerts a pull effect, drawing feature representations towards a predefined central point.
Conversely, for disease-positive bags that are primarily composed of negative tiles occasionally
interspersed with positive ones, the AD-loss initiates a push-pull effect. Results on two
histopathology datasets showcase our model's ability to boost classification performance by
minimising false negatives.
Citation
2025
DOI
Source Title
Publisher
Institute of Cancer Research (University Of London)
ISSN
eISSN
Collections
Research Team
Dynamical Cell Systems