Peripapillary Atrophy Segmentation with Boundary Guidance

Peripapillary atrophy (PPA) is a clinical finding that reflects atrophy of the retinal layer and retinal pigment epithelium. It is very important to segment PPA area as it indicates the progress of eye diseases such as myopia and glaucoma, while it is a challenging task to segment PPA due to the irregular and ambiguous boundaries. In this paper, a boundary guidance deep learning method is introduced to segment PPA area to obtain precise shape. We propose a boundary guidance block together with a contour loss function to improve the PPA segmentation performance on boundaries. Our approach is evaluated on a clinical dataset. The F1-score, IOU and Hausdorff distance of our method performance is 80.06%, 67.29%, 5.4934 respectively. Compared with other methods, our method achieves the best performance both qualitatively and quantitatively. Our proposed method can work well on retinal images with narrow PPA even with small training set.