Spatial alignment of MR brain images with an atlas

Sponsor:

Air Force Research Laboratories; Kettering Medical Center

Investigators:

Faculty: A. Goshtasby, M. Satter
Student: L. Ding, L. Zagorchev

A software system has been developed that spatially aligns an MR brain image with an atlas. The brain atlas, which is provided by Ron Kikinis, M.D., Harvard Medical School, comes with a volumetric MR image and a segmented label image. The steps of the alignment are as follows:

  1. Rigid registration: The MR atlas is rigidly registered with the given MR image using mutual information as the similarity measure. The details of this can be found in "Volume image registration by template matching" by L. Ding, A. Goshtasby, and M. Satter, Image and Vision Computing, vol. 19, 2001, pp. 821-832. The MR atlas is shown in the middle column of Fig. 2 and the segmented label atlas is shown in the middle column of Fig. 1. The given MR image is shown in the left column in Figs. 1 and 2. The right column in Figs. 1 and 2 show the overlaid images before the registration process. The result of the rigid registration is shown in Fig. 3. After this first step, the atlas image is resampled to a new image: atlas1.
  2. Nonrigid global registration: The rigid registration is examined and if necessary, the images are nonlinearly registered using thin-plate spline as the transformation function and a number homologous points in the given image and in atlas1. Since the rigid registration uses volumetric templates to achieve the matching, the centers of corresponding templates are used as homologous points. The homologous points are examined and if necessary, they are repositioned to improve the correspondence accuracy. From the correspondences the coefficients of a 3-D thin-plate spline are determine to transform atlas1 to register with the given MR image. This step transforms atlas1 to a new image: atlas2. The result of this step is shown in Fig. 4.
  3. Nonrigid local registration (coarse): Zero-crossing edges are determined in the given image and in atlas2 using a rather large Gaussian smoother. Weak edges are removed and correspondence is established between remaining edges. Edge correspondences are used as homologous points and coefficients of a locally sensitive transformation function are determined to transform atlas2 to overlay the given image. This step transforms atlas2: atlas3. The result of this step is shown in Fig. 5. There are some excessive deformations in this image. This is believed to be due to inaccurate edge correspondences. The correspondence process is currently being refined for a more accurate registration.
  4. Nonrigid local registration (fine): Edges are determined in the given image and in atlas3 using a rather small Gaussian smoother and after removing the weak edges, correspondence is established between the remaining edges. The correspondences are used to find a locally sensitive transformation function that resamples atlas3 to a new image (atlas4). Figure 6 shows the final result obtained in this manner.

At each step, as the MR atlas is transformed, the accompanying label image is transformed also. The user will be able to switch between MR and label images when overlaying the atlas with the given image. Overlaying of the label atlas with given MR image is shown in Fig. 7. When the label image is overlaid with the MR image, the user by moving the cursor on the MR image can read the atlas labels and thus determine the anatomic labels of voxels in the MR image.

The following figures show the flow of the registration process.

Fig. 1. (left) The given MR image. (middle) Segmented label atlas. (right) Overlaying of the left and middle column images.

Fig. 2. (left) The given MR image. (middle) MR atlas image. (right) Overlaying of the left and middle column images.

Fig. 3. Registration of the given MR and the MR atlas by rigid transformation.

Fig. 4. Registration using thin-plate splines.

Fig. 5. Local coarse deformation to register the atlas and the MR image.

Fig. 6. Final registration result.

Fig. 7. Switching the MR atlas with segmented label image. By clicking at a point in the given MR image, from the correspondence between the MR atlas and given MR image and the correspondence between the MR atlas and the labeled atlas, the anatomic label of the point in the MR image will be determined and printed out.


[Intelligent Systems Laboratory] [WSU Home Page] [CSE Department Home Page]

For more information contact A. Goshtasby (agoshtas@cs.wright.edu).

Last modified: 7/28/03