In the field of medical image analysis, supervised deep learning methods have become mainstream but still face significant challenges. Firstly, they require substantial annotated data, which is costly and time-consuming to obtain. Secondly, there is a class imbalance issue, especially with rare diseases that have limited images and diverse lesion types, making accurate modeling difficult. To address these issues, we propose an anomaly detection model based on knowledge distillation. Our approach leverages the differences in intermediate layer features between a pre-trained teacher network and a student network trained on normal images through knowledge distillation to detect anomalies. We incorporate a composite awareness module to learn both local regional and global spatial information effectively. Experiments conducted on three medical datasets prove our approach performs competitive anomaly detection results compared to existing methods for medical images.