Item Details

title

Theory and design of interpretable weakly supervised learning methods: application to breast cancer histopathology

author

Su, Ziyu

abstract

Histopathology is a fundamental tool for the diagnosis and prognosis of various diseases, including breast cancer [1]. In recent years, whole-slide imaging (WSI) technology has facilitated the analysis and annotation of histopathology images, providing opportunities for developing computer-aided diagnosis methods, particularly deep learning-based methods [2]. In addition, deep learning-based WSI analysis methods have advanced significantly during the last decade with multiple successful applications [3-7]. In general, these advancements offer strong indications that deep learning will play a crucial role as a necessary tool for histopathological analysis in clinical and biomedical domains.WSIs have two distinctive properties: large image size and difficulty in tissue-level annotation, making it challenging for regular deep learning models to process. In response to these challenges, researchers commonly employ weakly supervised learning methods that can process the WSI with only slide-level diagnostic labels. Multiple Instance Learning (MIL) stands out as one of the most popular paradigms within the weakly supervised learning family. Despite the success of existing MIL models, difficulties arise in certain WSI analysis tasks, particularly in addressing the disparity between slide-level labels and regions within the WSIs. This discrepancy often leads MIL models into a noisy training issue, as some input signals are irrelevant or even opposite to their labels [8]. To tackle this challenge, we propose a series of novel MIL methods designed to guide the model's focus on partially important regions within the WSI. Our approaches aim to mitigate the disparity between WSI inputs and slide-level labels. To validate our proposed methods, we apply them to two significant breast cancer histopathology problems: breast cancer lymph node metastasis (BCLNM) identification and Oncotype-DX (ODX) recurrence risk prediction. These tasks represent typical MIL scenarios. In the BCLNM task, predictive information is clearly defined but extremely small on the WSIs, while in the ODX task, predictive information is scattered across the WSIs but cannot be manually defined. In our experiments, the proposed methods exhibited excellent performance and continued to improve as we upgraded our methods. Importantly, the methods demonstrated great interpretability, enhancing their reliability for potential use in the clinical setting. Through these innovations, we envision the emergence of more accurate and clinically interpretable models, offering significant advancements in the diagnosis and prognosis of breast cancer.

subject

Breast Cancer

Computational Pathology

Deep Learning

Weakly Supervised Learning

Whole Slide Image

contributor

Gurcan, Metin N. (advisor)

Weaver, Ashley A. (committee member)

Tozbikian, Gary H. (committee member)

Lee, Sang Jin (committee member)

Gayzik, F. Scott (committee member)

date

2025-03-12T08:36:48Z (accessioned)

2025 (issued)

degree

Biomedical Engineering (discipline)

embargo

2025-09-11 (terms)

2025-09-11 (liftdate)

identifier

http://hdl.handle.net/10339/110322 (uri)

language

en (iso)

publisher

Wake Forest University

type

Dissertation