Abstract
Understanding the decision-making process of black-box neural networks is crucial for safe use of AI in high-stakes medical tasks such as histopathology. We present Adaptive Example Selection (AES), a prototype-based explainable AI framework that improves interpretability of deep learning models for mitosis detection. AES retrieves a sparse set of supporting and contradicting real-world prototype images to locally approximate the model's confidence surface with high fidelity ([Formula: see text]). The framework is integrated with a robust Faster R-CNN detector that demonstrates strong cross-tumor performance, for example achieving an F1-score of 0.84 on the Canine Cutaneous Mast Cell Tumor dataset. AES generates concise, case-specific explanations that faithfully capture local decision boundaries while linking predictions to interpretable exemplars. This enables clinicians to visualize model reasoning, assess uncertainty, and conduct contrastive analyses. Unlike prior methods focused on discrete class predictions, AES shows how similarity to mitotic and non-mitotic prototypes shapes graded confidence, enhancing transparency, trust, and practical adoption of AI-assisted mitosis detection in cancer diagnostics.