Enhanced Macular Telangiectasia Type 2 Detection: Leveraging Self-Supervised Learning and Ensemble Models.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Macular telangiectasia (MacTel) type 2 is a retinal disease affecting central vision.
• Accurate detection of MacTel type 2 is crucial for timely diagnosis and management.
• Optical Coherence Tomography (OCT) is a key imaging modality for retinal diseases.

Purpose of the Study:

• To develop and evaluate an ensemble-based deep learning model for detecting MacTel type 2 on OCT images.
• To assess the interpretability of the deep learning model in identifying MacTel type 2.
• To compare the model's performance against human expert graders.

Main Methods:

• Retrospective analysis of 5200 OCT scans from MacTel Registry and University of Washington.
• Training individual classification models using supervised and self-supervised learning (SSL).
• Ensembling individual models and evaluating performance using AUROC, AUPRC, accuracy, sensitivity, and specificity.

Main Results:

• The ensemble model achieved an AUROC of 0.972 and AUPRC of 0.967, comparable to human experts.
• High accuracy (91.7%), sensitivity (0.905), and specificity (0.925) were observed, despite limited labeled training data (10%).
• Individual models did not reach the performance levels of the ensemble.

Conclusions:

• Ensemble deep learning models, particularly when combined with SSL, offer accurate and interpretable detection of MacTel type 2.
• SSL effectively leverages unlabeled data, proving beneficial for rare disease detection with limited datasets.
• This approach shows promise for improving MacTel type 2 diagnosis and patient care.