Catching magnetic resonance imaging outliers in artificial intelligence-supported radiotherapy workflows: unsupervised detection and localization of image anomalies using deep learning

A fully automated, unsupervised anomaly-detection framework for pelvic and brain MRI has been developed, utilizing a two-stage architecture trained on public datasets such as LUND-PROBE and fastMRI. The framework compresses MRI slices into tokens and models the distribution of normal tokens, achieving AUCs of 0.97 for pelvic MRI and 0.81 for brain MRI in detecting anomalies. This approach enhances MRI quality control in radiotherapy workflows, providing interpretable visualizations of anomalies that could affect AI-driven clinical tasks.