Counterfactual explainability is an emerging class of xAI methods that aims to enhance interpretability by providing alternative scenarios to model predictions. These methods are particularly promising in medical imaging, where black-box deep learning models can be difficult to trust without intuitive explanations. In digital histopathology, counterfactual explanations could help pathologists understand why a model classified a tissue sample as malignant by generating alternative image versions that might have led to a different diagnosis.

This thesis explores the use of generative models for counterfactual explainability in digital histopathology, focusing on their ability to generate realistic and clinically relevant modifications of histopathological images. The goal is to determine how models such as variational autoencoders (VAEs) or generative adversarial networks (GANs) can be applied to generate counterfactual images that provide valuable diagnostic insights while preserving essential histological structures.

The student will conduct a literature review of generative counterfactual methods, select and implement a suitable approach, and apply it to explain the RationAI group's model trained to detect prostate cancer. All development and evaluation will be conducted within the RationAI group's cloud infrastructure.