Bilinear models for inter- and intra-patient variation of the prostate.

We propose bilinear models for capturing and effectively decoupling the expected shape variations of an organ both across the patient population and within a specific patient. Bilinear models have been successfully introduced in other areas of computer vision, but they have rarely been used in medical imaging applications. Our particular interest is in modeling the shape variation of the prostate for potential use in radiation therapy treatment planning. Using a dataset of 204 prostate shapes contoured from CT imagery of 12 different patients, we build bilinear models and show that they can fit both training and testing shapes accurately. We also show how the bilinear model can adapt to a new patient using only a few example shapes, producing a patient-specific model that also reflects expected content variation learnt from a broader population. Finally, we evaluate the training and testing projection error, adaptation performance and image segmentation accuracy of the bilinear model compared to linear principal component analysis and hierarchical point distribution models with the same number of parameters.