Date of Award

10-4-2023

Publication Type

Thesis

Degree Name

M.A.Sc.

Department

Electrical and Computer Engineering

Keywords

BYOL;Contrastive Learning;Domain Adaptation;Mammogram images;Self-Supervised Learning;SimCLR

Supervisor

Jonathan Wu

Rights

info:eu-repo/semantics/embargoedAccess

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Abstract

The most common cancer diagnosed worldwide is breast cancer and early detection is essential for reducing mortality. The best standard for early detection of breast cancer is digital mammography, which can aid physicians in treating the illness when it is still curable. However, inaccurate mammography diagnoses are frequent and can cause patients to undergo unnecessary examinations and therapies. This study aims to explore deep-learning techniques that can be utilized to implement and train a model to identify breast cancer cases in mammograms. Current deep learning-based diagnostic techniques are hindered by two fundamental issues: the expensive and time-consuming task of data annotation, and the lack of being adaptable to new or different data domains. This thesis uses cutting-edge deep learning methods to address these issues. We use self-supervised learning techniques like Bootstrap Your Own Latent (BYOL) and Simple Framework for Contrastive Learning of Visual Representations (SimCLR) to address the problem of scarce annotated data. These methods are contrastive learning methods which are a subset of self-supervised learning and have recently emerged as a crucial component for learning visual representations. These methods significantly reduce the need for labeled data while maintaining high model performance. Contrastive learning is still substantially unknown in the context of domain adaptation. These techniques have faced difficulties in their practical implementations due to issues with biased datasets that cause domain shifts. Domain shift occurs when distributions of data across domains have differences. In this research, we propose a technique to combine adversarial and contrastive learning to address the domain shift problem. By successfully reducing domain disparities, this strategy increases the model’s adaptability and robustness in a variety of clinical scenarios. Overall, our contributions provide an approach for early breast cancer detection that is more effective and flexible and has the potential to have significant effects on medical imaging diagnostics.

Share

COinS