The Use of Robust Local Hausdorff Distances in Accuracy Assessment for Image Alignment of Brain MRI

We present and implement an error estimation protocol in the Insight Toolkit (ITK) for assessing the accuracy of image alignment. We base this error estimation on a robust version of the Hausdorff Distance (HD) metric applied to the recovered edges of the images. The robust modifications we introduce to the HD metric significantly reduce the amount of outliers in the local distance error estimation. We evaluate the accuracy of our protocol on synthetically deformed images. We provide the source code and datasets to reproduce this evaluation. The proposed method is shown to improve error assessment when it is compared with conventional HD methods. This approach has many applications including local estimation and visual assessment of registration error and registration parameter selection.