Progressive Optimization of HydraLA-Net for Microaneurysm Segmentation

Abstract

Microaneurysms are the earliest detectable sign of diabetic retinopathy, yet automated segmentation remains challenging due to their small size, low contrast, and severe class imbalance in fundus images. In this work, we extend the Lesion-Aware Network (LA-Net) with class-specific prediction heads to reduce gradient competition during training. We conduct an experimental study on preprocessing techniques, including CLAHE variants, and imbalance-aware loss functions using a progressive optimization strategy across three public datasets: IDRiD, DDR, and TJDR. Results demonstrate improved microaneurysm segmentation while maintaining competitive performance on other lesion classes, providing a practical framework for enhancing early diabetic retinopathy detection. A full implementation is available at https://github.com/jessicayuan1/fundus-image-segmentation.