This thesis presents a novel way of integrating shape prior information into a level set based segmentation scheme. It utilizes the eigenimages of the signed-distance functions of the training shapes and confines the segmentation to statistically allowable shapes while minimizing the Chan-Vese functional via gradient descent. Implemented under the level set framework, the resulting algorithm can handle topological changes very well and is robust to noise and initial contour location due to the prior shape information being integrated. Meanwhile, the compactness of the eigenimage representation overcomes the "curse of dimensionality problem" existing for one-dimensional principal component analysis. We demonstrate this technique by applying it to several synthetic and real images.