Overview
Anomaly detection models identify images that deviate from learned normal patterns. These tools flag unusual cases for further review by clinicians. They are useful when labeled abnormal data are scarce.
Techniques
Autoencoders and generative models learn normal image distributions and detect outliers. Thresholding and uncertainty estimation guide flagging decisions. Model calibration is important to limit false positives.
Clinical Use Cases
Anomaly detection aids in screening for rare pathologies and quality control. It can detect unexpected device malfunctions or acquisition artifacts. Integration with workflows ensures timely human assessment.
Limitations
Models may flag benign variants as anomalies leading to unnecessary workup. Continuous refinement and clinician feedback reduce false alerts. Validation across populations ensures generalizability.