Explainability has increasingly become a core requirement for intelligent medical device software and AI models. Nevertheless, current medical AI technologies suffer from the 'interpretability gap', despite tremendous efforts for enhancing explainability. Here we propose class-association manifold learning, a generative AI technology to enhance explanation of black-box medical AI models. We adopt manifold learning techniques to efficiently decouples common decision-related patterns into a low-dimensional subspace against individual backgrounds, while preserving near-perfect diagnostic decision accuracy. The extracted pattern knowledge is further used to enable intentional modifications on arbitrary samples to visualize global diagnosis rules. We carry out experiments on an extensive set of medical AI tasks, showing that our method not only achieves much higher accuracy in explaining the behavior of medical AI models, but also helps with extracting medical-compliant knowledge that are unknown during model training, thus providing a potential means of assisting medical knowledge discovery.