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

Purpose: The purpose of this study was to develop a robust and general purpose artificial intelligence (AI) system that allows the identification of retinal optical coherence tomography (OCT) volumes with pathomorphological manifestations not present in normal eyes in screening programs and large retrospective studies. Methods: An unsupervised anomaly detection deep learning approach for the screening of retinal OCTs with any pathomorphological manifestations via Teacher-Student knowledge distillation is developed. The system is trained with only normal cases without any additional manual labeling. At test time, it scores how anomalous a sample is and produces localized anomaly maps with regions of interest in a B-scan. Fovea-centered OCT scans acquired with Spectralis (Heidelberg Engineering) were considered. A total of 3358 patients were used for development and testing. The detection performance was evaluated in a large data cohort with different pathologies including diabetic macular edema (DME) and the multiple stages of age-related macular degeneration (AMD) and on external public datasets with various disease biomarkers. Results: The volume-wise anomaly detection receiver operating characteristic (ROC) area under the curve (AUC) was 0.94 ± 0.05 in the test set. Pathological B-scan detection on external datasets varied between 0.81 and 0.87 AUC. Qualitatively, the derived anomaly maps pointed toward diagnostically relevant regions. The behavior of the system across the datasets was similar and consistent. Conclusions: Anomaly detection constitutes a valid complement to supervised systems aimed at improving the success of vision preservation and eye care, and is an important step toward more efficient and generalizable screening tools. Translational Relevance: Deep learning approaches can enable an automated and objective screening of a wide range of pathological retinal conditions that deviate from normal appearance.