Semi-supervised Segmentation Using Multiple Segmentation Hypotheses from a Single Atlas

A semi-supervised segmentation method using a single atlas is presented in this paper. Traditional atlas-based segmentation suffers from either a strong bias towards the selected atlas or the need for manual effort to create multiple atlas images. Similar to semi-supervised learning in computer vision, we study a method which exploits information contained in a set of unlabelled images by mutually registering them nonrigidly and propagating the single atlas segmentation over multiple such registration paths to each target. These multiple segmentation hypotheses are then fused by local weighting based on registration similarity. Our results on two datasets of different anatomies and image modalities, corpus callosum MR and mandible CT images, show a significant improvement in segmentation accuracy compared to traditional single atlas based segmentation. We also show that the bias towards the selected atlas is minimized using our method. Additionally, we devise a method for the selection of intermediate targets used for propagation, in order to reduce the number of necessary inter-target registrations without loss of final segmentation accuracy.