Addressing Class Imbalance in Diabetic Retinopathy Segmentation: A Weighted Ensemble Approach with U-Net, U- Net++, and DuckNet

Diabetic Retinopathy (DR) is a prevalent condition among diabetic patients, posing a significant risk of blindness if left untreated. Early detection of DR significantly improves the chances of successful treatment. The segmentation of DR lesions in Ultra-wide optical coherence tomography angiography (UW-OCTA) fundus images presents a complex challenge due to variations in lesion sizes and shapes. Traditional diagnostic methods by ophthalmologists are time-consuming. This study addresses lesion identification in an unbalanced dataset using deep learning techniques. The customized architectures such as U-Net, U-Net++, and DuckNet are parallelly trained on the dataset, and their performance is analyzed. Furthermore, an ensemble approach is implemented to enhance overall results, resulting in an ensembled model achieving a mean Intersection over Union of 0.52003. The mean Dice Similarity Coefficient, reaching 0.63269, demonstrates the highest accuracy for the Diabetic Retinopathy Analysis Challenge (DRAC) 2022 dataset, comprising UW-OCTA images of the retina.