Image Registration and Fusion: Task Group 132

what is the difference between image registration and image fusion

1. Image registration is the process of computing the geometrical transformation between two data sets (aligning them based on similarities).

   1. In other words: registration occurs when two or more separate imaging studies are assigned to a common coordinate system.

2. Once this is done, a fusion can be performed which consists of presenting the data sets together in a blended image.

3. This concept is commonly misinterpreted.  When the physicist or dosimetrist is figuring out the best overlay that two image sets can share, then they are doing a registration, not a fusion.  The fusion comes when they present the registration to the physician.  Thus, a good/bad registration results in a good/bad fusion.

What are some common reasons to perform image registration

1. Segmentation

2. adaptive Tx planning

3. image guided radiotherapy

4. image registration for response assessment

3 types of registrations

1. Rigid Registration - can include all translation and rotations but preserves the distances between all points in the image.

   1. Maximum of six degrees of freedom.

2. Affine Registration - includes all the rigid transformations but adds scaling, shearing, and plane reflection. Distances between points in the image are not maintained but parallel lines remain parallel.

   1. Maximum of 12 degrees of freedom.

3. Deformable Registration- includes all the affine transformations but allows spatially varying transformations.  Consider this like squishing play-doh into a mold.

   1. Maximum degrees of freedom of 3 times the number of voxels.

common metrics for evaluating a registration

1. Geometry-Based Metrics - This is the most common metric and makes uses of points, lines, or surfaces.  These features are then used as landmarks (and could be something like implanted fiducial markers).  The distances between points are compared.  The metric, represented by R in the following expression, may be calculated as for points p in studies A and B where N  is the number of points evaluated

2. Intensity-Based Metrics - This registration metric uses numerical grayscale information directly to measure how well two studies are registered.  Mutual information operates using intensity-based principles.  There are multiple metrics that can be used here we will examine a few:

   1. The sum squared difference formula compares intensities I in studies A and B for a number of voxels N. Requires images to be the same modality and have voxels with similar intensities

   2. Cross-Correlation - measures similarities in the image signal but can be sensitive to voxel intensity values if not normalized to itself.

   3. Mutual Information - useful for registering images from different modalities, this algorithm attempts to align voxels with Intensities that have a high probability of being present in both images.

List some quantitative methods for assessing registration accuracy

1. Target registration error (2-3 mm)

2. distance to agreement (2-3 mm)

3. Dice (above 0.8-0.9)

4. Jacobian (no neg values, values deviating as expected)

5. Consistency

Commissioning test reccs for a new image reg software

1. Phantom end-to-end tests:

   1. Important to have a phantom compatible with multiple imaging modalities such as CT and MR.

      1. Digital and physical phantoms are useful.

2. Correct image geometry:

   1. Check distances between known points on separate systems after transferring images.  

3. Correct image orientation:

   1. Make sure that if the CT scan was performed with a specific patient orientation (i.e. head first supine) then the registration software also displays the same.

   2. This should be performed for all clinically relevant patient orientations.