Detecting Anomalies in Retinal Diseases Using Generative, Discriminative, and Self-supervised Deep Learning.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Artificial intelligence (AI) systems for retinal diagnoses often require extensive training data for all disease variations.
• Anomaly detection offers a potential solution for AI systems lacking comprehensive training datasets for all retinal conditions and presentations.

Purpose of the Study:

• To investigate the efficacy of anomaly detection techniques in identifying retinal diseases, particularly when training data is limited.
• To explore the application of anomaly detection for general retinal disease screening and the identification of novel or rare retinal conditions.

Main Methods:

• Utilized a large dataset of high-resolution retinal fundus images from the EyePACS dataset.
• Trained 16 variants of anomaly detectors using only non-referable diabetic retinopathy images.
• Evaluated detector performance using metrics such as area under the receiver operating characteristic curve (AUC) and F1 score.

Main Results:

• The best-performing anomaly detector achieved an AUC of 0.808.
• This optimal performance was achieved using a self-supervised network embedding method.
• The system successfully identified referable diabetic retinopathy when trained solely on non-referable cases.

Conclusions:

• Anomaly detection techniques are valuable for retinal diagnostic systems trained with incomplete datasets.
• These methods can facilitate generalized retinal disease screening and the detection of rare or novel conditions.
• Anomaly detection holds potential for identifying unusual presentations of common retinal diseases.