Abstract <p>Anomaly detection is crucial in medical image analysis for identifying regions indicative of rare diseases. Due to the scarcity and high annotation cost of abnormal samples, unsupervised anomaly detection trained solely on normal data has become a key research focus. Current approaches are typically performed within a fixed latent probabilistic space, where the underlying distribution is learned from normal data and remains static during inference. However, these methods struggle to adequately capture the complex structures or multi-modal distributions of normal data, particularly in high-dimensional medical images, which often leads to blurred decision boundaries and hampers the effective differentiation of anomalous samples. To address this, we introduce a novel time-aware diffusion-enhanced reverse knowledge distillation framework that introduces the timestep in the diffusion process to elevate the original probabilistic space into a high-dimensional time-evolving dynamic latent space, and ultimately achieving both reconstruction and diffusion consistency. In addition, we propose a cross-consistency loss to achieve effective feature realignment between the teacher and student networks. Experiments on four different medical domains demonstrate the competitive performance of our method, highlighting its potential for advancing automated medical image anomaly detection.</p> Graphical abstract <p></p>