“…There are not many unsupervised anomaly detection baselines available for direct comparison due to three reasons: (i) no publicly released code [25,72,74,81], (ii) requiring weak/strong manual annotation [56,68,75,84], and (iii) relying on additional information from other modalities [17,46]. Therefore, we considered six major baselines for direct comparison: MemAE [13]-a Memory Matrix based method, Ganomaly [1]-a GAN-based method, f-AnoGAN [57]-the current state of the art for medical imaging, and CutPaste [37], PANDA [52], M-KD [56]the most recent unsupervised anomaly detection methods.…”