Unsupervised Anomaly Detection in Medical Imaging: A Survey of Methods, Challenges, and Future Directions

Unsupervised anomaly detection in medical imaging aims to automatically identify potential lesions that deviate from normal patterns in multimodal medical images without requiring annotations of abnormal samples, and is of great clinical value for early disease screening, unknown anomaly discovery, and label-scarce or open-set detection scenarios. Compared with industrial anomaly detection, medical images are characterized by complex anatomical structures, high semantic complexity of abnormalities, substantial inter-individual variability, and high annotation costs, which make the modeling and evaluation of related methods more challenging. This review systematically surveys unsupervised anomaly detection methods for medical imaging. By integrating task definitions, technological evolution, and clinical application needs, we comprehensively analyze 149 representative studies and 16 commonly used datasets collected from major academic databases. First, according to their core modeling paradigms, existing mainstream methods are categorized into four groups: image reconstruction-based methods, feature embedding-based methods, self-supervised learning-based methods, and foundation model-based methods. The technical characteristics, applicable scenarios, and inherent limitations of each category are then systematically discussed. Furthermore, from the perspectives of medical image structural properties and clinical application requirements, we summarize the key challenges currently faced by medical image unsupervised anomaly detection, including abnormal semantic modeling, the reliability of anomaly quantification, cross-center generalization capability, and evaluation protocols. Finally, future research directions, such as multi-task modeling, medical prior-guided learning, and multimodal fusion, are discussed in depth.