Robust 3D registration of CBCT images aggregating multiple estimates through random sampling

In this work, we propose a novel 3D rigid registration technique, by applying the Conventional Mutual Information based 3D Registration (CMIR) repeatedly in sampled data sets. Using the statistical distribution in the parameter space, we have considered mean, median, and mode of the distribution, and found that the median and mode provide reasonably good estimates. We call the method Robust Rigid Registration by Multiples Estimates in Sampled data points (R3MES). It provides higher accuracy than the CMIR technique. When the number of iterations of multiple estimates is kept low, the R3MES technique requires less computation, as it works in the sampled data set. The performance of the R3MES technique is better when the sampling rate is greater than 3%. We present a theoretical validation of this observation, considering uniform sampling with replacement. We also demonstrate its application for registering 3D CBCT image volumes of a colo-rectal cancer patient captured on different days. To demonstrate its general applicability, we present its performance on registering a pair of 3D brain MRI image volumes.