Group Equivariant Diffusion for Anomaly Detection in Computational Cytology

Researchers introduced a D4-equivariant diffusion model for cytology anomaly detection on whole-slide images.

The method uses symmetry in both training and inference to better reconstruct normal patches and rank abnormalities.

On bone marrow and peripheral blood smear datasets, it improved AUC, surfaced more abnormal cells in top-K results, and reduced score variance across rotated and flipped views.

The approach could make rare-cell screening more reliable when labeled abnormal samples are scarce.