Robust Registration of Statistical Shape Models for Unsupervised Pathology Annotation

Dana Rahbani, Andreas Morel-Forster, Dennis Madsen, Marcel Lüthi, Thomas Vetter

Abstract: We present a method to automatically label pathologies in volumetric medical data. Our solution makes use of a healthy statistical shape model to label pathologies in novel targets during model fitting. We achieve this using an EM algorithm: the E-step classifies surface points into pathological or healthy classes based on outliers in predicted correspondences, while the M-step performs probabilistic fitting of the statistical shape model to the healthy region. Our method is independent of pathology type or target anatomy, and can therefore be used for labeling different types of data. The method is able to detect pathologies with higher accuracy than standard robust detection algorithms, which we show using true positive rate and F1 scores. Furthermore, the method provides an estimate of the uncertainty of the synthesized label. The detection also directly improves surface reconstruction results, as shown by a decrease in the average and Hausdorff distances to ground truth. The method can be used for automated diagnosis or as a pre-processing step to accurately label large amounts of images.