Nadezhda Koriakina: Deep Learning and Explainable Artificial Intelligence for Biomedical Applications: Methods for Cytology-based Cancer Detection

Abstract

We live in an era of a rapidly evolving field of AI, where AI technologies have turned into an indispensable tool in a multitude of domains, including biomedical image analysis. Digital cytology, a field that deals with biomedical image data, could substantially benefit from AI technologies. AI techniques offer great potential to support medical experts in detecting diseases such as cancer, by alleviating the load on professionals and uncovering patterns that could go unnoticed by humans. However, AI algorithms come with ethical considerations and potential hazards that require attention and management. Recognising this problem is especially important in applications that handle patient data, given the serious consequences that could arise from mistakes. Another existing problem is that patient data often pose unique challenges that add complexity to the development of AI algorithms intended for handling such information.

This thesis comprises four papers that include methodologies for image classification of data with challenging properties, such as scarcity of fine-grained labels and complex data composition. Importantly, we explore those AI methods that are capable of addressing the lack of interpretability and trust in AI. Two of the four papers in this thesis are dedicated to making feasible end-to-end training of an interpretable multiple instance learning method on datasets with a large volume of data per patient, e.g., cytology data. The research work presented in one of the other papers of this thesis is focused on applying interpretable AI methods to analyse real-world cytology data for cancer detection. Motivated by the shortage of publicly available datasets in digital cytology and the scarcity of fine-grained labels in our real-world digital image cytology data, we investigate the role of synthetic data in the analysis of AI methods. In the fourth paper, we explore the capabilities of AI methods to analyse data with the sparseness of information relevant to a studied condition. This research question is important to answer for cytology-based early cancer detection. 

Our findings indicate that while image cytology data analysis comes with challenges, AI methods can play an important role in assisting medical experts by providing information that might prove valuable to them.