A modified Mumford-Shah model with nonlinear statistical shape prior was proposed to segment the optic nerve head in fundus images of poor quality, very low contrast, obscurity due to blood vessels, and distinct inter-differences of individuals. Firstly, the narrow band level set of shape prior was mapped into its kernel space by a nonlinear kernel function. Then, principal component analysis (PCA) was performed in the kernel space so as to acquire its base vectors, and statistical shape prior was integrated into a Mumford-Shah model. The segmentation results of the color optic nerve head images of patients in different stages of glaucoma have showed that the proposed model is effective and practicable.