Transformer-based self-supervised learning of pixel- and frequency-domain features for DMI grading on OCTA images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic macular ischemia (DMI) is a severe complication of diabetic retinopathy, threatening vision.
• Optical coherence tomography angiography (OCTA) is a key tool for non-invasive DMI diagnosis.
• Challenges exist in obtaining labeled OCTA datasets for DMI grading.

Purpose of the Study:

• To develop a self-supervised learning framework for DMI grading using OCTA images.
• To leverage both pixel- and frequency-domain features for enhanced DMI analysis.
• To improve the accuracy and efficiency of DMI progression diagnosis.

Main Methods:

• A self-supervised learning framework utilizing pixel-level and global filter transformers.
• Extraction of aggregated features from OCTA images via contrastive pretext tasks.
• Validation on three distinct retinal disease benchmarks.

Main Results:

• The proposed framework demonstrated superior performance over existing state-of-the-art methods.
• The system showed significant capabilities in recognizing DMI lesion areas.
• The approach holds promise for clinical diagnostic assistance in DMI.

Conclusions:

• Self-supervised learning offers a viable solution for DMI grading with limited labeled data.
• The framework effectively utilizes OCTA imaging for accurate DMI assessment.
• This technology has the potential to support ophthalmologists in diagnosing and managing DMI.