A new TVST study has introduced an unsupervised deep learning system capable of detecting anomalies in retinal OCT scans – without requiring any labeled pathological data. Developed by researchers at the Medical University of Vienna, the AI model leverages knowledge distillation to distinguish between healthy and abnormal retinal structures, offering a promising tool for automated screening of a diverse range of eye diseases.
The method is based on a teacher-student neural network framework. The “teacher” model, pretrained on natural images, guides a smaller “student” model, which learns to replicate the teacher’s outputs, but only using normal retinal OCT scans. At test time, deviations between the two models’ responses highlight anomalous regions, generating both anomaly scores and heatmaps that point to abnormal structures in the retina.
The system was trained and tested on a large in-house dataset of 3,247 OCT volumes, including a wide range of retinal diseases, such as diabetic macular edema (DME), age-related macular degeneration (AMD), Stargardt disease, and retinal vein occlusion (RVO). An average area under the curve (AUC) of 0.94 was achieved for detecting anomalous volumes, with particularly strong performance in diseases involving retinal deformation or fluid accumulation.
When tested on external public datasets (RETOUCH and Kermany), the model also performed well, with AUCs ranging from 0.81 to 0.87 at the B-scan level. Importantly, anomaly scores correlated strongly with disease severity, and the AI-generated maps successfully highlighted diagnostically relevant areas in the scans.
The authors believe this system could significantly reduce the workload of clinicians by better identifying slices of diagnostic relevance. The system’s explainability – via heatmaps that visualize abnormalities – also enhances clinical trust and integration. Unlike traditional AI systems that require thousands of labeled examples, this unsupervised approach only needs healthy data, making it scalable and adaptable for real-world settings. And while the system doesn’t yet identify specific diseases, and struggles with some large fluid regions, it represents a major step toward efficient, general-purpose OCT screening tools.
