The integrity and accuracy of image registration is dependent on the spatial accuracy of the original images. Scanners that supply images for image registration must be calibrated so that their voxel dimensions are known. Many scanners have a specified spatial accuracy no better than 1% and errors as high as 5% can occur. A 5% error on a 250 mm field of view corresponds to an error of 12.5 mm over the field of view, which will be unacceptable for most registration applications. Older CT scanners can have errors in axial dimensions due to inaccuracies in monitoring bed movement. Also, CT gantry tilt can be erroneously reported in image headers. A five-degree error will result in a shear of the data resulting in more than 20 mm error over a 250 mm field of view.
Scanner distortions can be categorized by those that distort the spatial integrity of the images and those that produce intensity shading effects. Geometric distortion is more likely in MR imaging. Distortion arises from gradient field non-linearity and from magnetic field inho-mogeneity. The former results from imperfections in gradient coil design but is static and can be measured and corrected. In practice, most modern MRI scanners will incorporate such a correction. Magnetic field inho-mogeneity arises from imperfections in the high Bo field, eddy currents induced by the switching gradients and spatial variations in the magnetic susceptibility within the imaged volume. Metal objects within or close to the field of view can result in severe distortions making image registration impossible. Smaller errors can be reduced by double acquisitions with inverted gradients  or phase unwarping methods . The former reportedly reduces errors by 30-40% while the latter reduced errors from 3.7 mm to 1.1 mm. Ramsey and Oliver  reported, in a recent study on modern MRI scanners, that linear distortions ranged between 0 and 2 mm. Modern MRI scanners are much less susceptible to geometric distortion than older machines.
Intensity distortion can arise from RF inhomogeneity in MRI, in particular with surface coils. Beam hardening effects in CT can produce intensity shading and photon scatter and incorrect attenuation correction can produce shading effects in PET. Highly attenuating objects will produce significant shading artifacts in PET and CT.
Objects of very different magnetic susceptibility to tissue can produce shading effects in MRI, and if made of metal will usually preclude acquisition of an MR image for registration. For this reason, patients with prostheses or surgical implants near the region to be imaged may not be suitable for image registration studies.
Mutual information-based registration has been shown to be reasonably immune to a gradual drop of intensity across the field of view. Studholme et al.  presented an adaptation of mutual information-based registration that improves robustness of MRI-PET registration of the pelvis based on intensity partitioning of the MR data, in cases where there is severe shading across the MRI field of view. Methods that rely on automated or semi-automated segmentation of surfaces may produce biased results in the presence of shading. Shading across an MR image can easily misplace a boundary by 1 or 2 mm over the field of view.
Ideally, image geometry should be measured during image acquisition, for example, by using markers in the patient-immobilization device. Great care must be taken that these markers do not themselves induce distortion or intensity artifacts. They are also used to measure geometric distortion on the periphery of the field of view, where it is most severe, running the risk of over-correction. Of more practical importance is careful quality assurance (QA) of scanners used for image registration. QA protocols should be similar to those implemented when images are used for high-precision radiotherapy or image-guided surgery. One method is to scan a simple geometric phantom regularly to check for distortions and scaling. Hill et al.  have proposed using an image registration method based on mutual information to measure scanner scaling and skew errors. A digital voxel model of the phantom is created and this is registered to the image volume using a full 12-degree of freedom affine transformation. This has been shown to be extremely robust and accurate and has been applied to PET, CT, and MRI scanner calibration. Scaling errors may be deduced in a similar manner by registration of a patient's CT scan and MRI scan and using a 9 degree of freedom registration (6 degrees of freedom rigid body plus scaling along the three ordinates).
Was this article helpful?