Unsupervised anomaly detection framework for retinal OCT imaging achieves strong cross-dataset generalization without disease-specific labels

Researchers developed an unsupervised anomaly detection method that learns normal retinal anatomy from healthy B-scans (two-dimensional cross-sectional images from Optical Coherence Tomography, a medical imaging technique) without requiring expert annotations of abnormalities, and detects anomalies via reconstruction discrepancies at both image and pixel levels.

The approach incorporates retinal layer-aware supervision and structured triplet learning to distinguish healthy from pathological representations. On the Kermany dataset, it achieved AUROC 0.799, outperforming VAE, VQVAE, VQGAN, and f-AnoGAN baselines; on the Srinivasan dataset, it achieved AUROC 0.884 with superior cross-dataset generalization.

On the external RETOUCH benchmark, unsupervised anomaly segmentation achieved Dice 0.200 and mIoU 0.117 scores, validating reproducibility across different institutions.