Résumé:
Diabetic Retinopathy (DR) is a frequent complication of diabetes mellitus that compromises
retinal function in more than 50% of type 2 diabetic patients. It occurs when the retina’s blood
vessels deteriorate. These altered vessels can dilate, leak fluid (plasma, lipids, and/or blood),
and even clog, leaving part of the retina without blood flow. All these phenomena that occur as
a result of diabetes can cause progressive damage to the structures of the eyeball, leading to
a severe reduction in vision and even, without appropriate treatment, to blindness in the
working age. In our work, we propose a framework for diabetic retinopathy detection based on
retinal lesions using advanced deep learning. We use the U-Mamba architecture for retinal
and blood vessel segmentation, achieving high F1 scores for various lesion types. Our Swin
Transformer-based classification model, incorporating lesion segmentation masks,
demonstrated exceptional performance across multiple datasets, with up to 97.75% accuracy
on the EyePACS dataset. This approach outperformed existing models across various
datasets, showing promise for clinical DR diagnosis.
