Comparison of different follow-up lung registration methods with and without segmentation

In modern multi slice CT scanners the increasing amount of data also increases the demand on image processing methods that assist the diagnosis. For the detection and classification of lung nodules in a follow up study it is very helpful to have the slices of a previous scan aligned with the slices of the current scan. This is a typical problem of image registration, for which different types of solutions exist. We investigated the accuracy and computation times of a rigid body, an affine, and a spline based elastic registration approach on the complete data set, and compared the results to a method where the registration was preceded by a segmentation of the lung volume. The registration quality was determined on a ground truth of previously determined lung nodule locations by measuring the average distance of corresponding nodules. It was found that an affine registration is slightly better than a rigid body registration, and that both are much faster than the elastic registration, which in turn showed the best registration quality. A good compromise was the affine registration on a previously segmented lung volume, which in total is not much slower than the registration without segmentation, but shows better alignment and higher robustness.