A comparison between adaptive kernel density estimation and Gaussian Mixture Regression for real-time tumour motion prediction from external surface motion

In this present study, tumour (3D) locations are predicted via external surface motion, extracted from abdomen/thoracic surface measurements that can be used to enhance dose targeting in external beam radiotherapy. Canonical Correlation Analysis (CCA) is applied to the surface and tumour motion data to maximise the correlation between them. This correlation is exploited for motion prediction [1]. Nine dynamic CT datasets were used to extract the surface and tumour motion and to create the Canonical Correlation model (CCM). Gaussian Mixture Regression (GMR) and Adaptive Kernel Density Estimation (AKDE) were trained on these nine datasets to predict the respiratory signal by updating the surface motion and CCM. A leave-one-out method was used to evaluate and compare the performance of GMR and AKDE in predicting the tumour motion.