Anomaly Detection and Biomarkers Localization in Retinal Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal diseases pose significant diagnostic challenges.
• Optical coherence tomography (OCT) provides high-resolution cross-sectional images of the retina.
• Developing automated methods for anomaly detection in OCT scans is crucial for timely diagnosis.

Purpose of the Study:

• To develop and evaluate novel artificial intelligence (AI) frameworks for anomaly detection and localization in retinal OCT scans.
• To assess the generalizability of self-supervised learning frameworks for identifying various retinal pathologies.
• To explore the potential of AI in uncovering novel biomarkers for retinal diseases.

Main Methods:

• Utilized four state-of-the-art self-supervised learning frameworks with pre-trained convolutional neural network (CNN) backbones.
• Applied frameworks to a combined dataset of publicly available (Kaggle) and local high-resolution retinal OCT scans.
• Evaluated anomaly detectors using metrics such as area under the receiver operating characteristic curve (ROC-AUC), F1 score, and accuracy.

Main Results:

• Trained on over 25,000 OCT scans, achieving high performance on test and validation sets.
• The best-performing framework achieved an ROC-AUC of 0.99, demonstrating excellent detection capabilities.
• Generated heat maps successfully localized anomalous regions, correlating with pathologies like choroidal neovascularization (CNV) and diabetic macular edema (DME).

Conclusions:

• Pre-trained feature extractors enable AI frameworks to generalize effectively to retinal OCT data.
• The developed frameworks demonstrate high accuracy in detecting and localizing retinal anomalies.
• These AI tools hold potential for clinical decision support, automated screening, and the discovery of new diagnostic biomarkers.