Regmentation: A New View of Image Segmentation and Registration

Abstract

Image segmentation and registration have been the two major areas of research in the medical imaging community for decades and still are. In the context of radiation oncology, segmentation and registration methods are widely used for target structure definition such as prostate or head and neck lymph node areas. In the past two years, 45% of all articles published in the most important medical imaging journals and conferences have presented either segmentation or registration methods. In the literature, both categories are treated rather separately even though they have much in common. Registration techniques are used to solve segmentation tasks (e.g. atlas based methods) and vice versa (e.g. segmentation of structures used in a landmark based registration). This article reviews the literature on image segmentation methods by introducing a novel taxonomy based on the amount of shape knowledge being incorporated in the segmentation process. Based on that, we argue that all global shape prior segmentation methods are identical to image registration methods and that such methods thus cannot be characterized as either image segmentation or registration methods. Therefore we propose a new class of methods that are able solve both segmentation and registration tasks. We call it regmentation. Quantified on a survey of the current state of the art medical imaging literature, it turns out that 25% of the methods are pure registration methods, 46% are pure segmentation methods and 29% are regmentation methods. The new view on image segmentation and registration provides a consistent taxonomy in this context and emphasizes the importance of regmentation in current medical image processing research and radiation oncology image-guided applications.