MIA_Registration

What does entropy?

-       Describes the distribution overlap

Does this image show a low or high entropy?

-       Low entropy à aligned

Does this image show a low or high entropy?

-       High entropy à not aligned

What is the problem if we using entropy?

If we just align black background, we have a minimum entropy à max alignment but the picture is not really aligned

What does mutual information

-       It not just considers the overlap area but also the content of the image

Which addition we have for the mutual information?

-       Additionally, to the joint histogram, we have an histogram of the image X and the image Y

What are the issues of mutual information?

-       Two images not the same size but the same information on the image (circle) Joint entropy favors the left transformation, MI favors the right transformation Normally not a big issue because one machine does always the same size of image à problem for studies in which different images from different hospitals are taken. Trick à normalize the mutual information

How we can optimize image registration?

Mostly based on classical optimization schemes à some form of gradient descent (E.g., Steepest gradient descent, Conjugate gradient descent) Means to find a potential minima!

What are some parameters of image registration optimization?

Maximum and minimum step size, stopping criteria, number of iterations Importance of Initialization (Manual, based on momentum, geometry of VOI) Number of samples for metric calculation (typically as a percentage of voxel number)

What is the aim of image registration evaluation?

-       Assess accuracy of th alignment

On what depends the complexity of the image registration evaluation?

Complexity of the evaluation depends on the type of registration

Name some image registration evaluation metrics:

Landmark based à typically uses RMSE (root-mean-square-error) metrics Hausdorff distance Uses of synthetic transformations Visualization of registration result à deformation of structured grid

How does the landmark-based evaluation works and what are the contras of this evaluation metric?

Landmark-based typically uses RMSE (root-mean-square-error) metrics Cons: Landmark annotation

How does the Hausdorff distance evaluation works and what are the contras of this evaluation metric?

-       doesn’t describe the entire transform

How does the deformation of structural grid evaluation works and what are the contras of this evaluation metric?

Deformation of structured grid à check for folds, unnatural local deformations Cons: difficult to check the entire 3D space

What is the main idea in nearest neighbor interpolation?

The idea is to use the pixel value of the data point or measurement which is closest to the current point. This is the fastest interpolation method but the resulting image may contain jagged edges.

What is the main idea in linear interpolation?

The idea is to survey the 2 closest pixels, then draw a line between them and designate a value along that line as the output pixel value.

What is the main idea in bilinear interpolation?

The idea is to survey the 4 closest pixels, then create a weighted average based on the nearness and brightness of the surveyed pixels and assign that value to the pixel in the output image. Use cubic convolution if a higher degree of accuracy is needed. However, with still images, the difference between images interpolated with this method and cubic convolution methods is usually undetectable. This interpolation usually supplies a much more viable alternative than the others.

For what is rigid registration appropriate?

Appropriate for: Brain (constrained by skull) Bone (neck, vertebrae)

For what is affine registration appropriate?

Appropriate for: If not all image acquisition parameters are known àunknown: voxel size, gantry tilt If scales changes are expected à growth, inter subject registration à Limited applicability expect as initialization for non-rigid registration