Review of Unsupervised Anomaly Detection Techniques for Medical Imaging

Unsupervised anomaly detection technology enabled the identification of out-of-distribution anomalous samples using only normal samples for training. Thus, it could be applied in the field of medical imaging for automated lesion detection and localization, playing a significant role in assisting disease screening, particularly in the diagnosis of rare diseases. Research progress on unsupervised anomaly detection in medical imaging. First, existing anomaly detection methods were categorized into four main types based on their core detection mechanisms, and the principles, characteristics, and representative works of each category were elaborated in detail. Second, commonly used public medical imaging datasets were summarized, and major anomaly scoring methods as well as benchmark evaluation metrics were introduced. Furthermore, through systematic quantitative experiments on multi-modal medical image datasets, the detection performance and computational efficiency of various representative methods were compared, and an in-depth analysis was provided based on the results. Finally, challenges in detection performance and clinical applicability were discussed, and future research directions were outlined.