Image Guided Neurosurgery System

Neurosurgery is an ideal application for image-guided techniques, by virtue of the high precision it requires, the need to visualize nearby tissue, the need for planning of optimal trajectories to target tissue, and the need to localize visually indistinguishable, but functionally different, tissue types. One method for aligning imagery to the patient is to use some form of extrinsic marker — a set of landmarks or other frame structures that attach to the patient prior to imaging, and that can then be used to establish a correspondence between position in the image and position on the patient. Examples include stereotactic frames, bone screws, and skin markers (e.g., [3,11,15,16, 18,19,23-26,28].

In other words, by placing markers on the patient prior to imaging, and keeping them rigidly attached through the completion of surgery, one obtains a coordinate frame visible in the imagery that directly supports transference of information to the surgical field of view. Stereotactic frames, though, not only are uncomfortable for the patient, but are cumbersome for the surgeon. They are limited to guidance along fixed paths and prevent access to some parts of the head. We would like to use a frameless system both for its simplicity and generality, and for its potential for use in other parts of the body. More recently, frameless stereotaxy systems have been pursued by many groups (e.g., [1, 2, 6, 21, 29, 35,]) and usually consist of two components: registration and tracking. We have added a third, initial, component to our system — reconstructed models of the patient's anatomy. The system's components are described next, with emphasis on the use of registration to align imagery with patient and surgeon's viewpoint.

The architecture of our image-guided surgery system (Fig. 1) supports frameless, nonfiducial, registration of medical imagery by matching surface data between patient and image model. The system consists of a portable cart (Fig. 2) containing a Sun UltraSPARC workstation and the hardware

FIGURE 1 Image-guided surgery system architecture.

to drive the laser scanner and Flashpoint tracking system. On top of the cart is mounted an articulated extendable arm to which we attach a bar housing the laser scanner and Flashpoint cameras. The three linear Flashpoint cameras are inside the bar. The laser is attached to one end of the bar, and a video camera to the other. The joint between the arm and scanning bar has three degrees of freedom to allow easy placement of the bar in desired configurations.

2.1 Imagery Subsystem

MRI is the prime imaging modality for the neurosurgery cases we support. The images are acquired prior to surgery with no need for special landmarking strategies. To use the imagery, it is important to create detailed models of the tissues being imaged. This means that we must segment the images: identify the type of tissue associated with each voxel (or volume element) in the imagery, and then create connected geometric models of the different types of tissue. Awide range of methods (e.g., [27,30,33,34,37]) have been applied to the segmentation problem. Classes of methods include statistical classifiers (e.g., [33,37]), which use variations in recorded tissue response to label individual elements of the medical scan, then extract surface boundaries of connected tissue regions to create structural models; deformable surface methods (e.g., [27,30]), which directly fit boundary models to delineations between adjacent tissue types; and atlas driven segmenters (e.g., [34]), which use generic models of standard anatomy to guide the labeling and segmentation of new scans.

Our current approach to segmentation uses an automated method to initially segment into major tissue classes while removing gain artifacts from the imagery [17,37], then uses operator-driven interactive tools to refine this segmentation. This latter step primarily relies on 3D visualization and data manipulation techniques to correct and refine the initial automated segmentation. The segmented tissue types include skin, used for registration, and internal structures such as white matter, gray matter, tumor, vessels, cerebrospinal fluid, and structures. These segmented structures are processed by the Marching Cube algorithm [20] to construct isosurfaces and to support surface rendering for visualization.

The structural models of patients constructed using such methods can be augmented with functional information. For example, functional MRI methods or transcranial magnetic stimulation methods (e.g., [9]) can be used to identify motor or sensory cortex. The key issue is then merging this data with the structural models, and to do this we use a particular type of registration method [7,8,38]. This approach uses stochastic sampling to find the registration that optimizes the mutual information between the two data sets. Optimizing mutual information makes the method insensitive to intensity differences between the two sensory modalities, and hence it can find the best alignment even if different anatomical features are

0 0

Post a comment