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

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Retinal optical coherence tomography (OCT) is crucial for diagnosing eye diseases.
• Current screening methods can be labor-intensive and may miss subtle pathological changes.
• Developing automated tools is essential for efficient and widespread eye care screening.

Purpose of the Study:

• To create a general-purpose artificial intelligence (AI) system for identifying pathological manifestations in retinal OCT volumes.
• To enable robust anomaly detection in screening programs and large retrospective studies.

Main Methods:

• An unsupervised deep learning anomaly detection approach using Teacher-Student knowledge distillation was developed.
• The system was trained exclusively on normal retinal OCT scans, requiring no manual pathological labeling.
• The system scores sample anomaly levels and generates localized anomaly maps on B-scans.

Main Results:

• The system achieved a volume-wise anomaly detection area under the curve (AUC) of 0.94 ± 0.05 on the test set.
• Pathological B-scan detection AUC on external datasets ranged from 0.81 to 0.87.
• Qualitative analysis confirmed that anomaly maps highlighted diagnostically relevant regions consistently across datasets.

Conclusions:

• Unsupervised anomaly detection is a valuable addition to supervised systems for enhancing vision preservation and eye care.
• This AI approach represents a significant step towards more efficient and generalizable retinal screening tools.
• Deep learning facilitates automated, objective screening for diverse pathological retinal conditions deviating from normal appearance.