Pedicle screw fixation is an established medical procedure for correcting congenital and acquired spinal deformities. The main challenge in this procedure is to insure that the screws are secured within the main axis of the pedicle, avoiding the spinal nerves. Accurate fixation of the screws requires insertion in the vertebral body through the axis of the pedicle. Therefore, during placement, the exact initial location and projection of the implanted screw is required. Pedicle screw placement utilizing image-guidance has become the standard of care in recent years. With this technology, the spine is exposed via surgical resection and then imaged (fluoroscopy or intra-operative imaging). The intra-operative image volume is then registered to the pre-surgical image volume to register the image-guidance system to the patient anatomy. Once patients have been registered to their image data, any device can be tracked in the space of their pre- or intra-operative image volumes.
One potential improvement on this technique is to perform the pedicle screw placement accurately without having to expose the spine. This may be accomplished if the pre-surgical image data provides all the information necessary to navigate in the space of the intact patient. In the case of spine surgery, this would entail the acquisition of tissue (muscle, nerve) and bone. Currently, high resolution MR images provide detailed images of tissue while CT images provide detailed bone images. This project has developed methodologies and software to interactively segment and register spinal MR and CT images to produce a “super image volume” exhibiting highly detailed images of both tissue and bone.
Registration of Spinal MR and CT Images Using the Mutual Information
Vertebral bodies represent rigid bodies and, therefore, individual vertebral bodies in MR and CT images can be registered by the rigid transformation. To register individual vertebral bodies in spinal MR and CT images, 1) image sub-volumes containing corresponding vertebral bodies are selected in the images and 2) the image sub-volumes are registered by the rigid transformation using the mutual information as the similarity measure. The process is demonstrated in Fig. 1. The same procedure may be used to register pre- and intra-operative CT images.
Fig. 1. (a) MR and CT spinal images of a patient before registration. Left column shows orthogonal cross sections of the MR image, middle column shows orthogonal cross sections of the CT image, and right column show overlaid MR and CT image volumes. (b) Approximate alignment of the images interactively. (c) Selection of sub-volumes containing the same vertebral body in MR and CT images. (d) Registration of the sub-volumes by rigid transformation using the mutual information.
Vertebral bodies have known shapes and sizes. Although variations of vertebral bodies exist from individual to individual, the variations are in size and are relatively small. Based on this information, models of a number of vertebral bodies are prepared by carefully delineating the vertebral bodies in a spinal CT image interactively. The created models are then used to segment vertebral bodies in a new spinal CT image automatically. This process is demonstrated in Fig. 2. By stacking the segmented slices a digital shape will be obtained, which can be triangulated for effective surface rendering. The triangular mesh is simplified to reduce the number of triangles and the simplified triangle mesh is subdivided to create a smooth surface at a desired resolution for rendering. This is demonstrated in Fig. 3.
Fig. 2. (a) – (d) A model vertebral body is created by interactively segmenting a spinal CT volume slice by slice.
Fig. 3. (a) A triangular mesh obtained by simplifying digital data obtained by interactive segmentation. (b) – (d) Representation of the model vertebral body at different resolutions obtained through mesh subdivision.
To segment a vertebral body in a new CT image, a model vertebral body is instantiated in the image interactively covering the area of interest. An energy minimizing surface is then initialized at the model surface and is allowed to deform based on edge information in the CT image. The model vertebral body will deform to take the shape of the vertebral body in the new CT image by minimizing its energy. This process is demonstrated in Fig. 4. Fig. 5 shows the obtained segmentation in 3-D.
Fig. 4. The process of taking a model vertebra and deforming it to resemble a vertebra in an image is based on an energy minimizing surface model. First, the approximate position and scale of the model in the CT volume is determined through interactive alignment. Then, a deformable mesh is initialized at the mesh representing the vertebral model. The energy of the deformable mesh is computed by inverse gradient magnitude at the surface points. Next, surface points are iteratively moved until minimum energy state is reached. The iterative process gradually deforms and repositions the model vertebra to take the shape and size of the new vertebra in the CT volume. The initial and final result are shown in (a) and (b).
(a) (b) (c)
Fig. 5. Different views of the deformed model delineating the new vertebral body in the CT volume.
Knowing the correspondence between CT and MR images obtained by image registration, the CT segmentation result can be transferred to the MR image as shown in Fig. 6. This process makes it possible to contain within the same MR volume details about both the soft tissues and the vertebral bodies.
Fig. 6. The surface of the segmented vertebra transferred to the MR image.
The segmentation process can be repeated to delineate all vertebral bodies of interest in the MR image. Fig. 7 shows segmentation of a sequence of vertebral bodies in this manner.
Fig. 7. The segmentation process is repeated to obtain all vertebral bodies.
MR images show soft-tissue details well while CT images show bone details well. By combining MR and CT images it becomes possible to see both soft-tissue and bone details well. A computational method is developed for the segmentation and registration of spinal MR and CT images for use in computer-assisted spinal surgery. In this project, first, the CT image is segmented to delineate individual vertebral bodies. Next, the MR and CR images are registered, and finally, the vertebral volumes delineated in the CT image are replaced with corresponding vertebral volumes in the MR image. This process enables viewing a single volumetric image with both soft-tissue and bone details, which can then become the input to an image-guided surgical system to provide effective navigation. By registering pre- and intra-operative CT images, it also becomes possible to evaluate progression of a surgery.