A shape prior extraction method based on two-dimensional principal component analysis (2DPCA)training is proposed for noisy image segmentation. Some ideal shapes are trained to obtain a group orthonormal projection vectors, and are then spanned to the 2DPCA space. The noisy image is projected onto this space, and least squares method is employed to find a dot nearest the projected dot in this space. The pre-image of this dot may not be one of the training shapes, but a linear combination of them. A pre-image strategy is then exploited to extract the shape prior. Experimental results show that the proposed segmentation method with shape prior is valid not only for noisy images, but also for images with occlusion and massing parts.