Interpretable AI methods for breast cancer whole slide image analysis

Histopathology is the diagnosis and study of diseases of the tissues which is undertaken by pathologists. For cancer diagnosis, they would traditionally examine tissue under the microscope and provide information to clinicians regarding cancer grade, type and potential response to treatment. However, with the introduction of digital pathology, a transition has begun away from traditional whole slide interpretation using microscopes. Instead, slides can now be viewed digitally, allowing for greater ease of use, rapid peer review and more efficient storage. Also, the availability of thousands of digitised histopathology slides facilitates the development of Artificial Intelligence based tools which can provide quick and accurate patient diagnostics. However, one drawback of these methods is, as their architectures have become more sophisticated, and performance increases, our ability to explain their decisions has decreased. The field of explainable AI attempts to remedy this by providing both models and tools for understanding the decision making process of AI solutions. In this thesis, methods for increasing the interpretability of Breast Cancer diagnostics models are explored. The main contributions are divided into three distinct chapters. In the first, a prototype based model, originally designed for fine-grained natural image classification, is used for interpretable breast cancer sub-typing and grading for the first time. The method achieves superior performance to less interpretable methods. The second contribution makes use of an attention based approach which allows for inspection of the most important image regions towards a classification. The approach is applied to a novel classification task which predicts the Nottingham Prognostic Index group. The approach is additionally extended to utilise both a patient's primary tumour site and lymph node image through late slide fusion, the latter of which is often neglected in automated diagnostic tools. The third contribution aims to combine the prototypical approach with the attention based approach. By making use of the prototype based approach, greater explainability can be provided to the attention mechanism. The methodology is used for both breast cancer grading and lung cancer sub-type prediction. Comparable performance is found with other state-of-the-art methods for lung cancer sub-typing and superior performance is found for breast cancer grading.