Statistical shape models (SSMs) are widely used for introducing shape priors in medical image analysis. However, building a SSM usually requires careful data acquisitions to gather training datasets with both suf?cient quality and enough shape variations. We present a robust framework to build reliable SSMs from a dataset with outliers and incomplete data. Our method is based on Point Distribution Models (PDMs) and makes use of recent advances in sparse optimisation methods to deal with erroneous correspondences. For validation, we apply the proposed approach to a dataset of 43 (including 24 corrupt) CT scans taken during routine clinical practice. We show that our method is able to improve the quality of the skull SSM in terms of generalization ability, specificity, compactness and robustness to missing data in comparison to standard and state-of-the-art algorithms.    «

Statistical shape models (SSMs) are widely used for introducing shape priors in medical image analysis. However, building a SSM usually requires careful data acquisitions to gather training datasets with both suf?cient quality and enough shape variations. We present a robust framework to build reliable SSMs from a dataset with outliers and incomplete data. Our method is based on Point Distribution Models (PDMs) and makes use of recent advances in sparse optimisation methods to deal with erroneou...    »