Head and Neck CT Atlases Alignment Based on Anatomical Priors Constraint

The contour delineation of radiotherapy target and organs at risk is an essential step for radiotherapy planning. Due to low accuracy and high labor intensity of manual delineation, machine learning-based automatic segmentation methods for CT images have become a research hotspot. Spatial alignment between atlases is an important preprocess step for the training of machine learning model and have tremendous consequences for the learning performance. However, because of low contrast of soft tissue in CT images, it is hard to obtain accurate atlases alignment by conventional methods, especially for the head and neck CT images including complex anatomies. In this paper, we propose a novel alignment method for head and neck CT atlases. Different from conventional methods which use information only from intensity images of atlases to be aligned and totally neglect the anatomical priors in atlases, our proposed method introduces anatomical priors to the optimisation procedure to improve atlases alignment performance. In our proposed method, CT intensity images and anatomical priors in atlases are jointly used to drive the initialization and evolution of deformation field that models the spatial correspondence between atlases. The contributions of our proposed approach are as follows: (1) a hybrid dissimilarity measurement with anatomical priors constraint, which uses not only information from intensity images but also anatomical priors of atlases, is developed to serve as the objective function of optimisation and drive the evolution of deformable model. (2) A construction method of B-splines initial deformable model based on anatomical priors is proposed to obtain better initial value for optimisation. Experiments for comparison with conventional atlases alignment method have been conducted on a public domain database for computational anatomy. Experimental results show that our approach can achieve much better results than conventional alignment methods.