Page 425 - Biomedical Engineering and Design Handbook Volume 2, Applications

COMPUTER-INTEGRATED SURGERY AND MEDICAL ROBOTICS 403

• Combine information of the same patient taken with different modalities, such as CT and MRI or
MRI and PET
• Combine information of the same patient before, during, and after surgery, such as preoperative
CT and intraoperative x-ray fluoroscopy, preoperative MRI and intraoperative video from a
microscope or an endoscope, or CT and x-rays from before and after surgery
• Create real-time virtual reality views of moving anatomy and surgical tools by matching preop-
erative models from CFT or MRI with intraoperative tracking data
• Perform a statistical study of patient data

Most CIS applications require more than one transformation to link two data sets, and thus have
more than one registration problem. For example, in the ROBODOC system, the preoperative plan
has to be registered to the intraoperative position of the bone so that the robot tip can machine the
desired canal shape in the planed position. To obtain this transformation, we must compute the trans-
formation from the bone coordinate system to the implanted fiducials, then from the fiducials to the
robot tip, to the robot coordinate system, and then to the cut volume. The series of mathematical
transformations that align one data set with another is called the registration chain.
The registration task is in fact not one but many different problems. There are great differences
on technical approaches, depending on the type of data to be matched, the anatomy involved, and the
clinical and technical requirements of the procedure. There is a vast body of literature on registra-
tion, which is comprehensively surveyed in Refs. 12 and 13 and can be classified according to the
following characteristics:
• Modalities. Refer to the sources from which data are acquired, for example, x-ray, CT, MRI, PET,
video, tracker. The combinations can be unimodal (same data source) or multimodal (different data
sources), which can be two images, an image to a model, or an image to a patient (tracker data).
• Dimensionality. Refers to the spatial and temporal dimensions of the two data sets to be matched
(two- or three-dimensional, static or time varying). The registration dimensionality can be static
2D/2D (x-ray images), 2D/3D (ultrasound to MRI), 3D/3D (PET to MRI) or time varying, such
as digital subtraction angiography (DSA).
• Registration basis. Refers to the image features that will be used to establish the alignment. These
can be extrinsic registration objects, such as a stereotactic frame or fiducial markers, or intrinsic,
such as anatomical landmarks, anatomical contours, or pixel intensity values.
• Nature and domain of mathematical transformation. Refers to the type of mathematical transfor-
mation that is used to perform the alignment. The transformation can be rigid, affine, projective,
or generally curved (deformable registration), and can be applied to parts of the image (local) or
to the entire image (global).
• Solution method. Refers to how the transformation is computed. This can include direct solutions
when an analytic solution or an appropriate approximation is found, or iterative solutions, where
there is a search and numerical optimization methods are used.
• Type of interaction. Refers to the type of input that the user has to supply. The registration is inter-
active when it is performed entirely by the user, automatic when no user intervention is required,
or semiautomatic when the user supplies an initial estimate, helps in the data segmentation, or
steers the algorithm by accepting or rejecting possible solutions.
• Subject. Refers to the patient source from which the images are taken; it can be the same patient
(intrasubject), two different patients (intersubject), or a patient and an atlas.
• Anatomy. Refers to the anatomy being imaged. This can be the head (brain, skull, teeth, nasal
cavities), the thorax (heart, breast, ribs), the abdomen (kidney, liver, intestines), the pelvis and the
perineum, or the limbs (femur, tibia, humerus, hand).
The main steps of registration algorithms are summarized in Table 14.2. Before attempting to
match the datasets, each data set should be corrected for distortions so that the errors resulting from