Unsupervised Anomaly Segmentation

Unsupervised anomaly segmentation aims to automatically identify unusual patterns in data, such as lesions in medical images, without requiring labeled examples of anomalies during training. Current research focuses on improving the accuracy and robustness of methods based on autoencoders, normalizing flows, diffusion models, and transformers, often incorporating techniques like iterative masking, cyclic modality translation, and adversarial learning to refine anomaly detection and reduce false positives. This field is crucial for applications like medical image analysis, where labeled data is scarce and efficient anomaly detection is vital for improved diagnosis and treatment. The development of more accurate and reliable unsupervised methods holds significant potential for advancing various scientific and practical domains.

Papers