Self-Supervised Representation Learning for Data-Efficient DRIL Classification in OCT Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Disorganization of the Retinal Inner Layers (DRIL) is a key biomarker for Diabetic Macular Edema (DME).
• Limited expert-annotated data hinders development of automated DRIL detection models.
• Domain variations in real-world images pose challenges for pre-trained models.

Purpose of the Study:

• To develop an efficient, self-supervised learning framework for DRIL detection.
• To overcome limitations of small, annotated datasets in medical imaging.
• To improve automated analysis of retinal abnormalities.

Main Methods:

• Proposed a novel two-stage, self-supervised learning framework.
• Utilized a large unlabeled Optical Coherence Tomography (OCT) dataset (108,309 images) for pre-training.
• Introduced spatial Bootstrap Your Own Latent (BYOL) with a hybrid spatial aware loss function.
• Fine-tuned the model on a small annotated OCT dataset (823 images) for DRIL classification.

Main Results:

• Achieved a high accuracy of 99.39% for DRIL detection.
• The proposed method significantly outperformed direct transfer learning models pre-trained on ImageNet.
• Demonstrated successful adaptation of learned representations for a specific pathological detection task.

Conclusions:

• Domain-specific self-supervised learning is highly effective for rare retinal pathologies.
• This framework addresses the challenge of limited annotated data in medical AI.
• The approach shows promise for enhancing diagnostic capabilities in ophthalmology.