Unsupervised deep generative models for anomaly detection in neuroimaging: A systematic scoping review

Abstract

Unsupervised anomaly detection (UAD) based on deep generative modelling has been increasingly investigated for identifying pathological brain abnormalities without requiring voxel-level annotations. By learning the distribution of healthy anatomy and generating pseudo-healthy reconstructions, these methods aim to localise deviations without requiring pathology-specific examples during training. Despite rapid methodological development – from autoencoders and variational autoencoders to generative adversarial networks and diffusion-based models – a structured synthesis of their application in structural neuroimaging is lacking. We conducted a PRISMA-ScR-guided scoping review of studies published or made available online between January 2018 and the search date (17 December 2025) that applied unsupervised deep generative models to anomaly detection in brain MRI (and, less frequently, CT). Thirty-three studies met the inclusion criteria. Methods were categorised by architectural family, and reported performance was synthesised across major pathology groups, with segmentation (Dice) and detection metrics (AUROC, AUPRC) disaggregated by evaluation level (voxel, slice, subject). For transparency, we also summarised dataset characteristics, dimensionality (2D vs. 3D), and thresholding strategies. Reported performance varied substantially across studies and pathologies. Dice scores generally ranged from approximately 0.30 to 0.75, while voxel-level AUROC values spanned roughly 0.58 to 0.97. Performance was typically higher for large focal lesions, such as brain tumours, than for small or diffuse abnormalities including white-matter hyperintensities. Although performance ranges overlapped substantially across architectural families, differences in evaluation level, thresholding strategy, dataset composition, and other experimental conditions preclude reliable attribution of cross-study performance differences to pathology-related factors or architectural design. Overall, unsupervised generative approaches show potential for anomaly localisation without requiring pathology-specific examples during training. However, methodological heterogeneity, limited external validation, and sensitivity to dataset characteristics remain major barriers to clinical translation. Emerging paradigms – including anatomy-aware modelling, diffusion-based frameworks, and alternative normative evaluation metrics – seek to improve robustness and support clinical translation.