Anomaly Detection in Retinal OCT Images with Deep Learning-Based Knowledge Distillation

AI Generated Summary*

An unsupervised deep learning system flags OCT scans deviating from normal anatomy via reverse knowledge distillation: a student network trained only on healthy B-scans reproduces a pretrained teacher's features, and mismatches yield an anomaly score plus an approximate localization map. The in-house cohort held 3247 volumes from 2713 patients covering intermediate and neovascular AMD, geographic atrophy, diabetic macular edema, Stargardt disease, retinal vein occlusion and central serous chorioretinopathy, with 176 normal eyes for training. Volume-level AUC was 0.94 (SD 0.05). On external sets, B-scan detection reached 0.81 (RETOUCH) and 0.87 (Kermany); RETOUCH positive predictive value averaged 0.90 but negative predictive value only 0.55. Map scores tracked lesion area (Spearman 0.78), though large fluid pockets were under-highlighted.

*This summary was generated by AI and is published unedited. Oku does not alter these summaries. It may contain errors or omissions and is provided for general informational purposes only. Oku does not guarantee its accuracy, completeness, or reliability. For authoritative information, please refer to the original, peer-reviewed article.