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

Diabetic macular ischemia (DMI) is a critical vision-threatening complication of diabetic retinopathy. While optical coherence tomography angiography (OCTA) enables non-invasive DMI progression diagnosis, acquiring labeled datasets remains challenging. This study proposes a self-supervised learning framework for DMI grading that leverages both pixel- and frequency-domain features. By using a pixel-level transformer and global filter transformer, the approach extracts aggregated features from OCTA images through contrastive pretext tasks. Experimental validation on three retinal disease benchmarks demonstrates superior performance compared to state-of-the-art methods, with the framework showing promising capabilities in lesion area recognition and potential clinical diagnostic assistance.