SU‐E‐T‐276: Learning Statistical Correlations of Deformation for Adaptive Radiation Therapy of Prostate Cancer

Purpose: The goal of this study is to develop a fast registration algorithm for aligning the planning image onto the treatment image of the same patient for adaptive radiation therapy of the prostate cancer. Methods: Our method is implemented by learning the statistical correlation of the deformations between prostate boundaries and internal regions from a population of training patients, as well as the online‐collected treatment images of the current patients. With the learned statistical correlation, the estimated boundary deformations can be used to rapidly predict the regional correspondences between prostates in the planning and treatment images. Initially, the population‐based correlation is used to predict the regional correspondences when the number of treatment images from the current patient is small. When more treatment images are obtained, the patient‐specific statistical correlation can take a more important role to reflect shape changes of prostate during the treatment. Eventually, only the patient‐specific statistical correlation is used to guide the regional correspondence prediction, once a sufficient number of treatment images are acquired from the current patient. Results: To demonstrate the performance of our method, we first evaluate its registration accuracy by comparing the deformation field predicted by our method with the deformation field estimated by TPS on serial CT images of 24 patients. The predictive error on the voxels around the prostate boundary is 0.38 mm by our method. It takes 24.5 seconds to register two prostate images, with comparable performance as TPS which needs 6.7 minutes for registering the same images. Conclusions: A new fast registration method has been presented to align the planning image onto each treatment image of a patient during radiotherapy. It can achieve similar registration accuracy as TPS, but performs much faster than TPS.