Brain Atrophy - Processing Options

Having set up the Brain Atrophy tool's input images as detailed in the Introduction, you can now set up the other processing options:

• . This tells the Brain Atrophy tool to use mutual information as the cost function when affine registering the template image to the T1-weighted input images. If the registration of the prior probability images fails, then it is worth trying the alternative cost function which is used for the initial affine registration. (The usual cost function is normalised correlation).

• . When selected, this tells Jim to apply extra regularisation when diffeomorphically registering the template image to the T₁-weighted input images. If the registration of the prior probability images results in severe local distortions, then it is worth trying to do the registration with extra regularisation.

• . This sets the strength of the Markov Random Field regularisation, which suppresses transitions in tissue between adjacent pixels during the segmentation. Selecting a higher value will lead to more uniform class probabilities, at the expense of being able to capture true changes in tissue class for small regions. You should normally not need to change the setting from its default value of 5.

• The type of lesion segmentation. These options are only available if you have set the number of image contrasts to 2, and are using both a T₁-weighted image and a FLAIR image as inputs (since the FLAIR image is needed for lesion segmentation).

  .

  • No lesion segmentation. Select this option if you want to use the FLAIR image to improve tissue classification, but there are no white-matter FLAIR hyperintensities that could cause misclassification.

  • Automatic lesion segmentation. White-matter FLAIR hyperintensities will be automatically segmented and outlined as regions-of-interest.
    You will also need to set the . The program is calibrated such that a setting of 0.5 gives satisfactory results in most cases. However, if you increase the lesion threshold fraction, the result will be a lower lesion volume; decreasing it leads to a larger lesion volume.

    If you trained the lesion segmentation using your own data, you will need to select , and set the file containing the trained image region statistics using: .

    You can later review the segmented lesions and make any corrections needed before running the classification again. If the lesion ROIs are already present (as for this second classification) no changes are made to them by the automatic lesion segmentation.

    Note: if you outline the lesion using ROIs on the input FLAIR images (saving the regions to a disk file with the default name for that image), then these will be transformed to the registered FLAIR image space, and used as lesion outlines instead of automatically segmenting the lesions.

  • Generate training data for lesion segmentation. Select this option if you wish to generate the data required for training the lesion classifier using your own data.

Several intermediate output images will be created in the selected folder. These are:

• Prior probability images. These are the template prior probability images, registered to the (co-registered) input T₁-weighted images, for the GM, WM, CSF and lateral ventricles.

• Pixel position information in template space.

• If you have more than one time-point, images of the Jacobian determinants will be created: an image for every pair of successive time-points. If, for example, you have set a base name of "Test", and there are two time-points, then an image called Test_Det_1_2 will be produced, which shows the Jacobian determinants for the deformation from the first time-point to the second.

• If you have more than one time-point, and are performing lesion segmentation, an image of the intensity difference in the FLAIR images for every pair of successive time-points will be created. If, for example, you have set a base name of "Test", there are two time-points, then an image called Test_D1_2 will be produced, which shows the change in intensity from the first time-point to the second. This is useful for identifying new and resolved lesions across time-points.

• Cross-sectional brain atrophy without lesions.
• Cross-sectional brain atrophy with lesions.
• Longitudinal brain atrophy without lesions.
• Longitudinal brain atrophy with lesions.