Anomaly detection finds statistical outliers in data. Machine learning based anomaly detection algorithms use a large number of normal examples to train an algorithm which detects what is normal (based on the training examples) and what is not normal. Anomaly detection algorithms can have features of both supervised and unsupervised learning, and is applicable to radiology as it is important to differentiate the normal from the abnormal. Some anomaly detection algorithms in radiology are implemented with unsupervised methods2,3 e.g. principal components analysis, and some have been implemented with supervised methods. Some basic anomaly detection algorithms for certain kinds of data e.g. measurement data derived from processed images, can be implemented with only classical statistical approaches.
Anomaly detection is especially useful when there are many different “types” of anomalies as it’s hard for any algorithm to learn from examples what different types of anomalies should look like and future anomalies may not show similarities to anomalies in any training data.
Unable to process the form. Check for errors and try again.