Diabetic retinopathy (DR) is the leading cause of blindness among working-age adults, making early detection crucial. Current automated systems classify DR severity but lack interpretability, limiting their clinical usefulness. This study presents a novel framework for training a deep learning model without human-labeled data to detect and segment four major DR lesions (microaneurysms, hemorrhage, hard exudate, and soft exudate) in retinal images. Using geometric rule-based algorithms, the model identifies high-probability regions, allowing for effective self-supervised training. Tested on public datasets, this approach significantly outperforms existing methods and is generalizable to other segmentation tasks.