OCT Image Generation for Subretinal Fluid Segmentation with Finite Element Structure Modeling

Deep learning has made significant progress in retinal edema segmentation. However, the limited availability of real data and the imbalance of pathological data have always hindered the performance of segmentation models. In this paper, we propose an image generation method based on finite element structural modeling (FESM), which ensures the pathological structure of synthetic data. This method fully utilizes the effective solution of finite element method in the process of generating OCT with subretinal fluid (SRF). Firstly, we decouple the structure of real data into SRF masks and layer segmentation, and use FESM to generate structure labels with different shapes. Then, a structure-guided latent diffusion generation model is proposed to learn the conditional distribution of images. We conducted extensive experiments on publicly available retinal segmentation datasets. The proposed method outperforms other generative models in terms of data quality and diversity. Furthermore, experimental results on downstream SRF segmentation tasks demonstrate that the data generated by our method can effectively improve the performance of many segmentation models. Our code is available at: https://github.com/duangduang0/FESM.