Advanced Retinal Lesion Segmentation via U-Net with Hybrid Focal–Dice Loss and Automated Ground Truth Generation

An early and accurate detection of retinal lesions is imperative to intercept the course of sight-threatening ailments, such as Diabetic Retinopathy (DR) or Age-related Macular Degeneration (AMD). Manual expert annotation of all such lesions would take a long time and would be subject to interobserver tendencies, especially in large screening projects. This work introduces an end-to-end deep learning pipeline for automated retinal lesion segmentation, tailored to datasets without available expert pixel-level reference annotations. The approach is specifically designed for our needs. A novel multi-stage automated ground truth mask generation method, based on colour space analysis, entropy filtering and morphological operations, and creating reliable pseudo-labels from raw retinal images. These pseudo-labels then serve as the training input for a U-Net architecture, a convolutional encoder–decoder architecture for biomedical image segmentation. To address the inherent class imbalance often encountered in medical imaging, we employ and thoroughly evaluate a novel hybrid loss function combining Focal Loss and Dice Loss. The proposed pipeline was rigorously evaluated on the ‘Eye Image Dataset’ from Kaggle, achieving a state-of-the-art segmentation performance with a Dice Similarity Coefficient of 0.932, Intersection over Union (IoU) of 0.865, Precision of 0.913, and Recall of 0.897. This work demonstrates the feasibility of achieving high-quality retinal lesion segmentation even in resource-constrained environments where extensive expert annotations are unavailable, thus paving the way for more accessible and scalable ophthalmological diagnostic tools.