Abstract. In this paper we propose an approach to generate a 4D statistical model of respiratory lung motion based on thoracic 4D CT data of different patients. A symmetric diffeomorphic intensity–based registration technique is used to estimate subject–specific motion models and to establish inter–subject correspondence. The statistics on the diffeomorphic transformations are computed using the Log–Euclidean framework. We present methods to adapt the genererated statistical 4D motion model to an unseen patient–specific lung geometry and to predict individual organ motion. The prediction is evaluated with respect to landmark and tumor motion. Mean absolute differences between model–based predicted landmark motion and corresponding breathing–induced landmark displacements as observed in the CT data sets are 3.3 ± 1.8 mm considering motion between end expiration to end inspiration, if lung dynamics are not impaired by lung disorders. The statistical respiratory motion model presented is capable of providing valuable prior knowledge in many fields of applications. We present two examples of possible applications in the fields of radiation therapy and image guided diagnosis. 1

Modeling of respiratory motion gains in importance within the field of radiation therapy of lung cancer patients. Current modeling approaches are usually confined to intra-patient registration of 3D image data representing the individual patient’s anatomy at different breathing phases. We propose an approach to generate a mean motion model of the lung based on thoracic 4D CT data of different patients to extend motion modeling capabilities. Our modeling process consists of two main parts: an intra–subject registration to generate subject–specific motion models and an inter–subject registration to combine these subject–specific motion models into a mean motion model. Further, we present methods to adapt the mean motion model to a patient-specific lung geometry. A first evaluation of the model was done by using the generated mean motion model to predict lung and tumor motion of individual patients and comparing the prediction quality to non–linear registration. Our results show that the average difference in prediction quality (measured by overlap coefficients) between non–linear registration and model–based prediction is approx. 10%. However, the patient–specific registration relies on individual 4D image data, whereas the model–based prediction was obtained without knowledge of the individual breathing dynamics. Results show that the model predicts motion patterns of individual patients generally well and we conclude from our results that such a model has the capability to provide valuable a-priori knowledge in many fields of applications. Categories and Subject Descriptors (according to ACM CCS): G.3 [Probability and Statistics]: Time series analysis 1.

journal homepage: www.elsevier.com/locate/media Location registration and recognition (LRR) for serial analysis of nodules

Abstract. Detailed analysis of breathing dynamics, as motivated by radiotherapy of lung tumors, requires accurate estimates of inner lung motion fields. We present an evaluation and comparison study of non-linear non-parametric intensity-based registration approaches to estimate these motion fields in 4D CT images. In order to cope with discontinuities in pleura and chest wall motion we restrict the registration by applying lung segmentation masks and evaluate the impact of masking on registration accuracy. Furthermore, we compare diffusive to elastic regularization and diffeomorphic to non-diffeomorphic implementations. Based on a data set of 10 patients we show that masking improves registration accuracy significantly. Moreover, neither elastic or diffusive regularization nor diffeomorphic versus non-diffeomorphic implementation influence the accuracy significantly. Thus, the method of choice depends on the application and requirements on motion field characteristics. 1