Towards transparent computer vision for deep learning in natural and histopathology imaging

Deep learning based artificial neural networks increasingly become the state of the art in more and more fields. The recent emergence of LLM based GenAI tools like ChatGPT has raised the global awareness for AI, and these tools have quickly become widely adopted. While very powerful, they lack transparency, tend to hallucinate and can pick up unwanted biases out of their training data. To address these problems, explainable AI (xAI) tools are being actively researched and developed, in order to make the models more transparent and spot biases in them.

This thesis focuses on xAI methods applied to computer vision models in natural and medical imaging, aiming to gain insight into those black-box neural networks. Current methods are so called saliency maps that show where the model is looking, visualization methods that show what the model is seeing and concept based methods that combine them both to give a global explanation of the behavior of the model. The contributions of this thesis are as follows. A method to modify existing models to generate smoother saliency maps. In the area of visualization based methods, we apply them to generative models resulting in a transparent model, as well as another method using a generative model to speedup the generation of visualizations by 120x compared to iterative gradient descent with comparable quality. In the area of concept based methods, we propose an automated test as well as a method to contrast any two classes, instead of explaining them in isolation. These contributions resulted in four peer reviewed publications and one preprint, as well as two presentations at medical conferences and one pending patent.