Install the "Simple Elastix Registration" Action

CamiTK provides an specific extension that uses the elastix toolbox directly in order to offer simple image registration.

Elastix is an open source software, based on the well-known Insight Segmentation and Registration Toolkit (ITK or Insight Tool Kit). Elastix consists of a collection of algorithms that are commonly used to solve medical image registration problems.

Elastix provides two applications:

• elastix is a program that computes a registration between a two images. It requires an input parameter file and an output directory (more about it later)
• transformix is a program that uses the resulting transformation computed by elastix to apply it to an input image. It requires an image, an input transformation file and an output directory (more about it later)

First you will need to install the CamiTK extension, then when you use the extension, you will need to specify the path to the elastix program.

The "Simple Elastix Registration" action generates the required elastix parameter files, launch the elastix program, and open the resulting registered image in CamiTK.

Please follow these installation steps:

• Step 1. Download the Windows extension archive.
• Step 2. Unzip simple-elastix-registration-win.zip anywhere you like (e.g., in your Downloads directory or C:\Temp).
• Step 3. Open camitk-imp, and then open the Edit → Preferences menu or press Ctrl+P.
• Step 4. Open the Action Extensions tab and click on the Add button:

• Step 5. Select the elastixregistration.dll file you just extracted. The "Simple Elastix Registration" action should now appear in the list, click on it to get its description:

• Step 6. On Windows download the latest elastix software version corresponding to your OS from the official elastix github pages (you should also read the legal stuff). On Linux, install the official elastix package.
• Step 7. Unzip the elastix-5.1.0-win64.zip anywhere you like (e.g., in your Downloads directory or C:\Temp).
• Step 8. On Windows, please make sure you have installed the latest supported version of the Microsoft Visual C++ Redistributable. Choose the one corresponding to your architecture (if you are on a laptop, it should be X64). Reboot after installation.

All set! You should be able to use elastix directly from CamiTK imp.

Objectives of this exercise

In this exercise, we will first artificially displace the T1 medical image volume by manually changing its position and orientation and then try to register it back to the initial non-displaced image.

Simple Elastix Registration parameters

But first, lets check the parameters that can be modified in the Simple Elastix Registration action. Later on, when you will trigger the action, the parameters will look like this:

The following parameters are available:

• Translation Initialization: whether or not the initial translation between images should be estimated as the distance between their centers.
• Transform Type: determines what type of deformations between the fixed and moving image is handled. There are 6 models available in elastix. Only three different types are available in this action: translation only, affine transformation (includes translation, rotation, scaling and shearing), and the nonrigid (i.e., elastic) B-spline transformation (uses cubic multidimensional B-spline polynomial control points). For more information, see section "2.6 Transform" of the elastix manual for more information.
• Similarity Metric: the similarity measure that is used to compare the voxel values and find the best match during the registration. There are 5 different similarity measures available in elastix. Only four are available in this action: Advanced Mean Squares Difference, Advanced Normalized Correlation, Advanced Mutual Information and Normalized Mutual Information. See section "2.3 Metrics" of the elastix manual for more information.

During this tutorial, we are going to explore the effects of this three parameters. The CamiTK action also has the following parameters:

• Result Name: this is automatically deduced from the choice of the three elastix parameters explained above. This name will be given to the image computed by elastix. You may change it if you wish.
• Moving Image: this is the image that elastix will try to move/deform to match the fixed image.
• Fixed image: this is the target of the registration. Elastix is trying to compute a mapping from the fixed image to the moving image, i.e., ${}^{moving} T_{fixed}$.
• Overwrite Results: if checked, the action will overwrites the previous elastix results when the same registration is rerun.

Testing Registration with a Small Translation

First we will try registering the T1 image with a slightly translated version of itself.

Testing Registration with small rotation

We will now try registering the T1 image with a slightly rotated version of itself.

Make it fail

We will now apply a bigger rotation to T1 so that there is no overlap between the new and original image to see what elastix can do.

And succeed again

Now we understand the basics of registration, lets try to use it to register something more useful in the case of the brain shift application.

In the registration directory, you will find another image: FLAIR_perop.mhd. We will consider that this image is acquired using an intra-operative MRI after the surgeon opened the skull to get access to the tumor.

As shown in the previous exercise, the registration will only converge if the image are not too far from each other. In this exercise, we consider that a first registration was done in order to pre-register the intra-operative image FLAIR_perop.mhd to the pre-operative image FLAIR.mhd

Initial affine registration

We are first going to register the two images using affine registration. The translation initialization should not be required here as the images were pre-registered.

From now on, we will consider that FLAIR_perop.mhd is the fixed image, and that FLAIR.mhd is the moving image. The image resulting from the registration is therefore going to be a transformed FLAIR.mhd that match a much as possible (and depending on the parameter we use) the intra-operative image FLAIR_perop.mhd.

We will now use the following parameters:

Changing metric

As you can see, the registration is not perfect:

• There is of course no deformation as the transform model is set to affine.
• By trying to register all the data, the minimization process of the affine transform model resulted in a scaling perturbation

We are now going to check the effect of the choice of a metric, and change from Mean Square Deviation to Mutual Information.

Using elastic registration

Lets try now to find the deformation component. For that we need to change the transform type to an elastic transform model. We are going to use a B-spline transformation.

Registering the pre-operative mask

Now that we have found a correct elastic registration, we are going to use it to deform the brain mask segmented on the pre-operative image.

This is an important step in build a navigation system. Using the transformation found by the elastic registration algorithm, we will be able to transfer all the planning the surgeon has done in the pre-operative images into the intra-operative space.

In the previous part of this tutorial, we have produced a mask of the brain tissue and a mesh to represent it in 3D. There could be other pre-processing stage done on the pre-operative image, when time is not as crucial as during the operation. For instance, another mask could have been segmented on the pre-operative images in order to delimit the brain contour (including for instance a security margin). Or the brain tissue mask could have been register to a complex brain atlas in order to label all the tissues and areas of the brain.

The per-operative segmentation is an important step and allows the surgeon to define her/his planning (e.g., planning to resect of the tumoral tissues). Unfortunately, as soon as the skull is opened to get access to the tumor, brain-shift phenomenon means that the tumor is not at the place it was relatively to the skull in the pre-operative images. All the pre-operative planning information have to be registered as well to the per-operative space.

This is basically what we are going to do in this exercise by using the previously found elastic registration transformation on the pre-operative segmented mask.

In order to apply a transformation, we are now going to use transformix, the second application provided by the elastix software.

We will now consider the following inputs:

• the mask mhd image file (called FLAIR_mask.mhd in the remaining of this exercise)
• the transformation file, produced by elastix. If you did not modify the default result name parameter, this file should be called Transform-FLAIR-to-FLAIR_perop-noinit-b_-nmi.txt on your computer.

Refine the registration using the deformed mask as the ROI

Another very useful thing that we can do using the registered mask is to use it to define a Region Of Interest for the registration algorithm. Instead of asking the whole image as the input for registration algorithm, we can now perform a second registration step and ask it to only focus on the region defined by the mask. This will produce a better registration of the brain tissue (but as you will see, it might produce an alien skull!)

To give the mask to the elastix application, we are now going to directly use the elastix binary on the command line. This means that we will first have to manually create a new subdirectory for elastix to write its results.