On-line 3Dmotion estimation using low resolution MRI

Abstract

Abstract.Image processing such as deformable image registration finds its way into radiotherapy as a means to track non-rigid anatomy. With the advent of magnetic resonance imaging (MRI) guided radiotherapy, intrafraction anatomy snapshots become technically feasible. magnetic resonance (MR) imaging provides the needed tissue signal for high-fidelity image registration. However, acquisitions, especially in 3D, take a considerable amount of time. Pushing towards real-time adaptive radiotherapy, MR imaging needs to be accelerated without degrading the quality of information.In this paper, we investigate the impact of image resolution on the quality of motion estimations. Potentially, spatially undersampled images yield comparable motion estimations. At the same time, their acquisition times would reduce greatly due to the sparser sampling. In order to substantiate this hypothesis, an exemplary 4D dataset of the abdomen is downsampled gradually. Subsequently, spatiotemporal deformations are extracted consistently using the same motion estimation for each downsampled dataset. Errors between the original and the respectively downsampled version are then evaluated.Compared to ground-truth, results show high similarity of deformations estimated from downsampled image data. Using a dataset with (2.5mm) 3 voxel size, deformation fields could be recovered well up to a downsampling factor of 2, i.e. (5mm) 3 . In a therapy guidance scenario MRI, imaging speed would accordingly increase approximately fourfold, with acceptable loss of estimated motion quality.

Intrafraction motion quantification and planning target volume margin determination of head-and-neck tumors using cine magnetic resonance imaging

“…Deformable image registration was used to quantify tumor with sub-voxel precision. Sub-voxel precision of the specific algorithm used in this work was demonstrated to detect deformations at approximately 1/3 of the pixel size [29]. Tumor motion was quantified over a 8 min period approximately 5 and 13 min after the patient entered the MR scanner.…”

Anatomically plausible models and quality assurance criteria for online mono- and multi-modal medical image registration

“…Performance of the original Horn-Schunck algorithm While the HSO has been shown to be a reliable performer by several independent studies (Ostergaard et al 2008, Zachiu et al 2015b, Glitzner et al 2015, Yang et al 2016, the assumption that a material point conserves its intensity as it moves, however, can also be a source of misregistrations (Ostergaard et al 2008, Martel et al 2007, Zachiu et al 2015b). In the current work, this is best illustrated for the MR -MR mono-modal registration in figure 2 and 4.…”

“…The smoothness regularization basically hypothesizes that neighboring points within an image usually move together (or at least in a similar fashion) and penalizes motion patterns such as shearing. The Horn-Schunck algorithm has been adopted successfully for medical image registration in several independent studies (Ostergaard et al 2008, Glitzner et al 2015, Yang et al 2016. Although the algorithm suggested by Horn & Schunck was originally not developed for medical image registration and thus does not specifically take into account the material properties of biological tissues, it has been shown to be a reliable performer for the monomodal image registration of soft-tissue structures.…”

Accelerating multi-modal image registration using a supervoxel-based variational framework

“…We also recall that it is common to accelerate the registration process by simply running the algorithm on down-sampled versions of the two input images (Glitzner et al 2015, Zachiu et al 2017c. The obtained deformation field is then upsampled at the original image resolution and the magnitude of each displacement vector is adjusted accordingly.…”

“…While employing various regularization terms such as smoothness, curvature or incompressibility, the computational demand of the above-mentioned multi-modal registration algorithms remains predominantly determined by the calculation of the data fidelity term. In the context of a clinical workflow, the computation time is a critical point and several minutes at least are typically required for the registration of 3D images of dimension 128 × 128 × 128 using above mentioned methods.A simple manner to accelerate the registration process consists in running the algorithms on a down-sampled version of the input images (Glitzner et al 2015). This comes, however at a cost in precision and accuracy of the motion estimates due to a loss of image features or occurrence of artifacts during the down-sampling process.…”