Evaluation of Plastimatch B-Spline Registration on the EMPIRE10 Data Set Gregory Sharp 1 , Marta Peroni 1,2,3 , Rui Li 1 , James Shackleford 4 , Nagarajan Kandasamy 4 1 Massachusetts General Hospital, Boston, MA 2 Politecnico di Milano, Milan, Italy 3 Massachusetts Institute of Technology, Cambridge, MA 4 Drexel University, Philadelphia, PA Abstract. In our open source software package “Plastimatch”, we pro- vide a B-spline based deformable image registration method with an efficient GPU and multicore implementation. We have participated in the EMPIRE10 grand challenge to evaluate our method on the task of registering a set of benchmark thoracic CT data sets. The results demon- strate that our method ranks 12 on the 34 methods evaluated. On the set of statistics we computed, we have shown that our registration methods can register the benchmark images at full resolution in 0.4 ∼ 5.7 minutes with good results based on the Dice and invertibility statistics. Keywords: Deformable Image Registration, B-spline 1 Introduction Deformable image registration is an important and challenging research area in medical imaging, and it has been successfully applied to various problems in the clinical setting. One such problem is pulmonary CT image registration as highlighted in this EMPIRE10 grand challenge. Due to the elastic nature of lung tissue deformations, deformable image reg- istration is needed to register a pair pulmonary CT images. We have developed a B-spline based deformable image registration method with efficient implemen- tation as part of our open source software package “Plastimatch” [1]. By using a grid alignment scheme in our method, we can significantly ac- celerate the B-spline interpolation and gradient computation, thus speeding up the registration process. The improved efficiency of the registration process is reported as running time on the benchmark dataset provided by EMPIRE10. It takes 0.4 ∼ 5.7 minutes to register to a pair of 3D CT benchmark images at full resolution. One of our goals to participate this grand challenge is see how accurate our method compared to other methods. Based on the evaluation metrics, we achieved an average ranking of 12 out of 34 methods submitted. Combined with our own evaluation metrics of Dice coefficients and root mean squared error (RMSE), our method is accurate and certainly demonstrate the potential of clinical applicability. However, as all these evaluation metrics are not conclusive