<p>Deep learning-based models have achieved high accuracy in histopathology and are a valuable support tool for medical specialists, helping them enhance diagnostic accuracy, obtain objective results, and reach faster diagnoses. However, their lack of transparency, or black-box nature, raises concerns as it reduces trust in the models’ predictions. Explainable Artificial Intelligence (XAI) provides tools to make black-box models more transparent and understandable for pathologists, thus improving reliability in critical tasks such as histopathological diagnosis. We present HistoXplaining in Style, a method that leverages a StyleGAN2-ADA based approach to create high-fidelity synthetic histopathology images as counterfactual explanations for image classification tasks. In particular, our method enables highly efficient generation of counterfactual images compared to state-of-the-art approaches, being 72 times faster, while maintaining high visual quality with an FID of 16.2. A remarkable finding is that the intermediate latent space of the generative model organizes tissue features in a semantically meaningful way, clearly separating benign from cancerous phenotypes. In a validation study involving 13 medical experts, we show the realism of our generated images and the clinical utility of the counterfactual explanations. Notably, 92.3% of experts rated them as “very” or “extremely useful” for understanding and trusting the AI’s predictions. Thus, by combining XAI counterfactual explanations and generative models, we provide a potential tool for improving the interpretability and trust of AI in histopathology.</p>