In image segmentation, the shape knowledge of the object may be used to guide the segmentation process. From a training set of representative shapes, a statistical model can be constructed and used to constrain the segmentation results. The shape space is usually constructed with tools such such as principal component analysis (PCA). However the main assumption of PCA that shapes lie a linear space might not hold for real world shape sets. Thus manifold learning techniques have been developed, such as Laplacian Eigenmaps and Diffusion Maps. Recently a framework for image segmentation based on non linear shape modeling has been proposed; still some challenges remain, such as the so-called out-of-sample extension and the preimage problems. This paper presents such a framework relying on Diffusion Maps to encode the shape variations of the training set, and graph cut for the segmentation part. Finally, some segmentation results are shown on a medical imaging application.