Multi scale self supervised learning for deep knowledge transfer in diabetic retinopathy grading.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a primary cause of vision loss, requiring timely and precise detection.
• Automated deep learning models face challenges with complex retinal images and insufficient labeled data.
• Traditional transfer learning methods often underperform in medical imaging due to domain discrepancies.

Purpose of the Study:

• To develop an advanced deep learning model for improved diabetic retinopathy grading.
• To overcome limitations of existing models in handling medical image complexity and data scarcity.
• To leverage self-supervised learning with a hybrid architecture for enhanced feature extraction.

Main Methods:

• Proposed a Multi-scale Self-Supervised Learning (MsSSL) model integrating Vision Transformers (ViTs) for global context and Convolutional Neural Networks (CNNs) with a Feature Pyramid Network (FPN) for multi-scale features.
• Employed a Deep Learner module to refine extracted features, enhancing spatial resolution and capturing both high-level and fine-grained details.
• Utilized domain-specific pretraining to adapt models to medical imaging data.

Main Results:

• The MsSSL model demonstrated significant improvements in diabetic retinopathy grading accuracy.
• Performance surpassed traditional deep learning and transfer learning methods.
• The study highlighted the effectiveness of combining global and multi-scale feature extraction for medical image analysis.

Conclusions:

• The MsSSL model offers a promising advancement for automated diabetic retinopathy detection and grading.
• Integrating ViTs and CNNs with FPN provides a robust framework for medical image analysis.
• Domain-specific pretraining and advanced model architectures are crucial for success in medical AI.