Evaluation of the feasibility of anomaly detection for dose management in PET examinations

Owing to the revision of the Medical Care Act in 2020, managing and recording radiation doses in PET-CT examinations have become mandatory. In this study, we investigated unsupervised anomaly detection methods as a potential solution to minimize input errors in dose recordings.

We analyzed data extracted from our database, including patient body weight, positron emission tomography (PET) dose, and dose length product (DLP). Several anomaly detection models, such as one-class support vector machine (OCSVM), Hotelling's T2 method, multivariate statistical process control (MSPC), isolation forest, and local outlier factor (LOF), were applied and compared. The dataset included 3509 entries for model training and 499 entries for evaluation. Anomalies that could be potential input errors were evaluated using metrics, such as precision, recall, F1 score, receiver operating characteristics-area under the curve (ROC-AUC), and precision-recall-AUC (PR-AUC).

We demonstrated that Hotelling's T2 method and MSPC's T² statistic outperformed other models, achieving a recall of 1.0 and AUCs of 1.0, effectively detecting input errors in radiation dose records. Furthermore, our findings suggest that unsupervised anomaly detection can not only identify input errors but also detect excessively high or low radiation doses, contributing to improved dose management in PET-CT examinations.

These findings suggest that unsupervised anomaly detection is a promising approach to improve the accuracy of dose management in PET-CT examinations, enhancing patient safety and compliance with regulatory standards.