RESPL-Net: A novel computer-aided lesion segmentation system for fundus images in retinopathy of prematurity

The diagnosis and treatment of retinopathy of prematurity (ROP) is crucial, yet there is limited research on computer-aided diagnosis of ROP that focuses on the pathological structure of fundus images in premature infants. It mainly faces two challenges. Firstly, labeling ROP images is costly, and there is currently no publicly available dataset containing segmentation images of lesions for the first three stages of ROP. Additionally, the lesion areas of ROP are small, and their features are less distinct compared to the background, so lesion knowledge is difficult to learn and represent. To address these issues, this paper first constructs a dataset for ROP lesion segmentation, named ROP-PS99, which includes lesions from the first three stages. The ROP-PS99 dataset serves provides pixel-level annotated images for lesion segmentation, supports the evaluation of computer-aided diagnostic methods for ROP. Meanwhile, we propose a new ROP lesion segmentation method, named RESPL-Net. In the RESPL-Net, we introduce a suitable pixel-level attention module (PL) that can bridge local lesion descriptors with their long-range dependencies of fundus images in ROP. Furthermore, we propose DFLoss, which addresses the issues of weak lesion texture, difficulty in distinguishing lesions from surrounding regions, and class imbalance in ROP detection. The experimental results demonstrate that the proposed ROP lesion segmentation method outperforms other baseline methods.