Building Shape Models from Lousy Data

Statistical shape models have gained widespread use in medical imageanalysis. In order for such models to be statistically meaningful, a large number ofdata sets have to be included. The number of available data sets is usually limitedand often the data is corrupted by imaging artifacts or missing information. Wepropose a method for building a statistical shape model from such ”lousy” datasets. The method works by identifying the corrupted parts of a shape as statisticaloutliers and excluding these parts from the model. Only the parts of a shape thatwere identified as outliers are discarded, while all the intact parts are includedin the model. The model building is then performed using the EM algorithm forprobabilistic principal component analysis, which allows for a principled wayto handle missing data. Our experiments on 2D synthetic and real 3D medicaldata sets confirm the feasibility of the approach. We show that it yields superiormodels compared to approaches using robust statistics, which only downweightthe influence of outliers.