Fast and efficient image registration based on gradient orientations of minimal uncertainty

There exist a wide variety of time sensitive contexts (e.g. image-guided neurosurgery (IGNS)), whereby image registration is required to be both fast and accurate if it is to be adopted clinically. Many sampling techniques have been proposed to speed up the registration process but these often come at the expense of accuracy (e.g. random). In this paper, we describe a fast and accurate multi-modal registration framework based on matching gradient orientations at locations of minimal gradient magnitude uncertainties in a coarse-to-fine manner. In the context of IGNS, the method was shown to perform with accuracies below 2mm using 2% of the total voxels when tested on the 14 cases of the publicly available BITE dataset [1]. For rigid registration between MRI and CT brain images on the RIRE dataset [2], the quantitative results demonstrate that the proposed approach can employ highly reduced sampling rates (e.g. 0.05% of the voxels in the image) while still yielding a median registration error inferior to 1mm [3]. In the context of the non-rigid registration of inter-patient MRI brain volumes, the proposed approach is evaluated with a publicly available dataset, and achieves comparable accuracy to the top performing methods but with only one sixth of the processing time [4]. While the results are promising, there are remaining challenges associated with existing sampling techniques, as well as limitations in the existing validation frameworks for registration.