Peripapillary atrophy (PPA) is a kind of abnormality that serves as a potential indicator for myopia and glaucoma. Accurate segmentation and measurement of PPA region can assist doctors in better diagnosing and monitoring fundus diseases. Current research for PPA segmentation achieves higher accuracy, but the shapes of some segmentation are unreasonable. Due to the relatively regular shape of PPA, if the prior shape information is integrated, the segmentation result will be more reasonable. In this paper, we first use a deep learning-based segmentation network to predict the PPA region coarsely, and a post-processing method with active shape constraint is proposed to refine this region. To obtain the shape descriptors, a novel automatic landmark sampling method is proposed, in which the auxiliary angle-finding method is employed to handle potential segmentation discontinuities in practical application. In active shape model training, a dictionary selection algorithm is proposed to automatically select the most representative shapes as training set to avoid overfitting. Extensive experiments on a clinical dataset demonstrate that our proposed method provides good qualitative and quantitative performance.