Mild Cognitive Impairment (MCI) is an important stage between Normal Control (NC) and Alzheimer's disease (AD), thus the early detection of MCI carries important clinical significance. Anomaly detection can identify abnormal regions in medical images and serve as a tool for early diagnosis and treatment of MCI patients. However, the absence of pixel-level labels in the MCI database poses a challenge to performing anomaly detection using supervised learning. Unsupervised learning can be utilized as an alternative approach to address the above-mentioned issues. In this paper, we propose an unsupervised anomaly detection framework based on the diffusion model for the early detection of MCI. The proposed framework learns the distribution of healthy data and guides the abnormal data to conform to the healthy distribution, thereby achieving "healing" and locating anomaly regions. Experimental results demonstrate that the proposed framework can effectively distinguish between healthy and abnormal data, and highlight anomaly regions associated with MCI, showing its practical significance in the early diagnosis of MCI.