Image registration

This article uses direct referencing.

Image registration is the process whereby features in one digital image are aligned with corresponding features of a reference image, usually referred to as the template. The most common purpose of this undertaking is to allow for quantitative comparisons between the two images.

In principal, any digital image can be registered onto any other, though useful comparisons can obviously only be made if there is sufficient correspondence between their features. This is typically the case if the same or similar objects (e.g., your nose and that of your mother) have been recorded using the same imaging modality but in some contexts, it is not uncommon to image the same object with different imaging modalities. In the case of clinical diagnostics, for instance, the options to image a patient's brain or knee would include at least the following:

1. Focus on structure: MRI, CT, X-ray, US, IR, digital photographs (e.g. of knee flexion, histological samples or light microscopic scenes);
2. Focus on function: PET, SPECT, fMRI, contrast-enhanced CT/MR/US, digital videos.

Of course, each modality may have subtypes reflecting different image acquisition parameters (e.g., the brain ventricles typically appear dark in T1-weighted MR images, bright in T2-weighted ones, and gray in MR images weighted by proton density), and it may be necessary to register images of different kinds, e.g. a T2-weighted image to a T1-weighted template.

A great number of approaches to image registration exist (Hill et al., 2001) but they all include the following basic steps (Crum et al., 2003):

1. specify a transformation type (e.g., rigid body, affine, elastic, fluid, B-spline, n-point support),
2. specify a similarity or error measure quantifying the quality of the match (e.g., least squares, correlation ratio),
3. specify an interpolation strategy (e.g., nearest neighbour, trilinear, sinc),
4. find the transformation parameters to optimise the similarity measure.