Application of Unsupervised Anomaly Detection on Time Series Medical Radiation Data

This paper examines the application of three unsupervised anomaly detection techniques on time series medical radiation data: Long-Short Term Memory (LSTM) autoencoder, Reconstruction Convolutional Autoencoder (RCAE) and Isolation Forest. An unlabelled time series radiotherapy dataset with artificially inserted anomalies was used to evaluate the performance of these models, with a goal to identify anomalous data in medical radiation which could signify dangerous spikes or inadequate dosages for a patient. The results indicates that the LSTM autoencoder model showed the best performance in accurately detecting anomalies within in the medical radiation dataset, due to its ability to capture long-term temporal dependencies. Meanwhile, RCAE performed the best on a separate dataset from the Numenta Anomaly Benchmark (NAB), suggesting its strength in handling datasets with structured or spatial patterns. The study highlights the importance of considering temporal relationships for anomaly detection in sequential data and provides insights into the strengths of different models for specific applications.