FreeLesion: Synthetic Image-Mask Pairs for Fundus Lesion Segmentation via Curriculum Learning and Feature-Loss Guided Filtering

Abstract

Fundus lesion segmentation plays a crucial role in the automated diagnosis of diabetic retinopathy. While recent advances in lesion segmentation networks have shown promising results, they are often constrained by limited pixel-level annotations, leading to suboptimal model training. In this paper, we introduce FreeLesion, a deep generative model-based approach for synthesizing image-mask pairs to mitigate the issue of data scarcity. Specifically, we employ a curriculum learning strategy that progressively enables the generative model to learn from easy to hard lesion categories. This approach enhances the model’s ability to generate smaller, more morphologically complex lesions. Additionally, we propose a feature-loss guided filtering method to remove noisy synthetic regions, improving the alignment between generated images and their masks. Experimental results demonstrate that our synthetic fundus images significantly boost lesion segmentation model performance, reaching new state-of-the-art performance. Notably, training the segmentation model using only images generated by our approach yields better performance than using real images alone (36.2 vs. 35.6 in mIoU on IDRiD dataset). Our code will be released at https://github.com/guomugong/FreeLesion.