Overcoming Data Scarcity in Retinal Lesion Segmentation via Instance Injection and Multi-Scale Ensembling

Automated segmentation of retinal lesions is a prerequisite for the early diagnosis of Diabetic Retinopathy (DR). However, the development of robust models is hindered by the extreme scarcity of pixel-wise annotations and the severe class imbalance between structural features and minute pathologies. In this work, we propose a data-efficient framework that improves upon the standard Swin-UNet architecture. We introduce a Coordinate-Attention Fusion module to recover spatial information lost during downsampling and implement a novel training pipeline integrating Copy-Paste Instance Injection and Multi-Scale Zoom Augmentation. To ensure statistical robustness, we employ a 5-Fold Cross-Validation Ensemble. Benchmarking against a standard ResNet34 U-Net baseline on the IDRiD dataset, our method demonstrates superior convergence. While the baseline suffers from oversegmentation (Mean IoU: 0.2346), our proposed framework achieves a Mean IoU of 0.3606, representing a 53.7% relative improvement. Notably, we achieve state-of-the-art recovery of challenging Soft Exudates (IoU 0.58 vs. Baseline 0.12).