Continuous Sampling in Mutual-Information Registration

Mutual information is a popular and widely used metric in retrospective image registration. This metric excels especially with multimodal data due to the minimal assumptions about the correspondence between the image intensities. In certain situations, the mutual-information metric is known to produce artifacts that rule out subsample registration accuracy. Various methods have been developed to mitigate these artifacts, including higher order kernels for smoother sampling of the metric. This study introduces a novel concept of continuous sampling to provide new insight into the mutual-information methods currently in use. In particular, the connection between the partial volume interpolation and the recently introduced higher order partial-volume-type kernels is revealed.