Deep Learning Framework Detects Anomalies in Pelvic and Brain MRI

A fully automated, unsupervised framework for anomaly detection in pelvic and brain MRI has been created by researchers, targeting weaknesses in AI-enhanced radiotherapy systems. This two-stage model utilized reference images from public datasets: LUND-PROBE for pelvic MRI and IXI, fastMRI, and fastMRI+ for brain MRI. Initially, MRI slices are transformed into discrete tokens, followed by modeling the distribution of normal tokens. Anomaly detection is achieved by integrating perceptual image differences with token-surprisal scores derived from negative log-likelihood. The research delves into deep learning-based anomaly detection for pelvic MRI, an area that has seen limited exploration, and assesses its potential for complete automation, aiming to identify out-of-distribution images that could disrupt clinical operations.