Explainable Fundus Image Curation and Lesion Detection in Diabetic Retinopathy

Fundus photos capture the structure of the retina along with abnormalities indicative of the stage of Diabetic Retinopathy. Artificial Intelligence (AI) can help clinicians identify these lesions, reducing manual workload, but models require high-quality annotated datasets. Due to the complexity of retinal structures, errors in image acquisition and lesion interpretation of manual annotators can occur. We proposed a quality-control framework that ensures that only high-standard data is used for AI training. First, an explainable feature-based classifier is used to filter inadequate images. The features are extracted using both image processing and contrastive learning. The images are then enhanced and subject to annotation, using deep-learning-based assistance. Lastly, the agreement between annotators based on derived formulas determines the usability of the annotations.