Internal motion prediction using kernel density estimation and general canonical correlation model

This paper presents preliminary work in developing a global correlation model between lung tumor respiratory motion and external surrogate motion in external beam radiotherapy. This involves using a combination of set of dynamic CT datasets to train a bivariate kernel density estimation model. Canonical correlation analysis (CCA) is used to parametrize the correlation between the external observation surrogate and the target region, in this case tumor temporal motion. Such an approach is non-invasive and non-ionizing, and minimizes the patient setup time. Preliminary results shows that the correlation coefficient for preliminary data is high, ranging between 0.87 and 0.99. Recasting the internal and external motion into eigenspace reveals the underlying correlation in an efficient and compact manner. A leave-one-out method was used to validate the proposed algorithm. The average error of tumor position was about 1.6 mm.