Atlas-based segmentation has become a standard paradigm for exploiting prior knowledge in medical image segmentation. In this paper, we propose a method to exploit both the robustness of global registration techniques and the accuracy of a local registration based on level set tracking. First, the atlas is globally put in correspondence with the patient image by an affine and an intensity-based non rigid registration. Based on this rough initialisation, the level set functions corresponding to particular objects of interest of the deformed atlas are used to segment the corresponding objects in the patient image. We propose a technique to derive a dense deformation field from the motion of these level set functions. This is particularly important when we want to infer the position of invisible structures like the brain sub-thalamic nuclei from the position of visible surrounding structures. This can also be advantageously exploited to register an atlas following a hierarchical approach. Results are shown on 2D synthetic images and 2D real images extracted from brain and prostate MR volumes and neck CT volumes.
I. IntroductionAtlas-based segmentation of medical images has become a standard paradigm for exploiting prior anatomical knowledge in image segmentation. Some of the most critical requirements of atlas-based segmentation, particularly in radiation therapy or neurosurgical planning, are the following: the contours of segmented structures have to be found as accurately as possible, while staying well smooth, the connectivity relationships between structures defined in the atlas have to be maintained through the registration / segmentation process and the segmentation of structures without visible edge, i.e. contours only defined with respect to adjacent structures, has to be possible. In the majority of the approaches proposed so far to register an atlas to a patient image, the objective of the transformation is to optimize some global intensity-based correspondence measure (like level-gray differences, regional correlation, or mutual information). Some recent algorithms like [1] combine global and more local intensity-based registration. This permits to improve the results while decreasing the computation time. However this is not yet sufficient to guaranty the desired quality of segmentation for the most demanding applications. Most of the time the only constraint used on the transformation is its smoothness, ensured for instance by a Gaussian filtering [2] or constraints between interpolation functions [3]. When at some places contours are not accurate enough, it is usual to globally or locally allow more elasticity to the deformation in order to obtain a more local deformation, with the risk of increasing the irregularity of the deformation field and thus of the contours, without necessarily obtaining the sought level of precision.To cope with this problem, additional constraints have to be included in these types of intensity-based registration algorithms. These constraints should permit to impos...