Explainable machine learning models to assist with cancer diagnosis

This thesis proposes a package of machine learning tools to assist in the classification process of cancers through medical imaging. Enhancing the interpretability of a given machine learning method is a key focus of the work. Firstly, a “patient like me” methodology using the Fisher Information Network is created to show a robust structure of clinical data from a neural network leading to new patient cases being classified in a visual way. Next this thesis studies the partial responses of a neural network to understand how changes in values of important variables affect the contribution towards the final prediction of the classifier. This attempts to reflect clinical thinking, where changes in variables would change the clinical outcome of a decision-making process. Finally, the thesis looks at multimodality data fusion to utilise as much of the abundance of available clinical data as possible. This work looks at the effectiveness of information with an approach that includes a feature selector. The three aspects of work have been assessed with publicly available clinical datasets to allow for clinical meaning to be ascertained. The thesis looks at potential clinical impact throughout and how the application of machine learning can be useful in a clinical setting, rather than only providing a classification output