Three dimensional multimodal image registration using implanted markers
Venkateswara R. Mandava · 1992
Our goal is to register arbitrarily oriented, multimodal, volume images of the human head by aligning a configuration of three or more fiducial points. Each point is the geometric center of a marker implanted in the skull. To compute the centers we have extended an earlier thresholding algorithm based on image statistics. The extension, which relies on qualitative knowledge of the histogram in the region surrounding and including the marker, permits the segmentation of regions of interest with multiple objects. We have simulated markers of varying size and varying intensity in CT, MRI T1-weighted, MRI T2-weighted, and MRI proton density images, segmented the simulated markers, computed the geometric centers, and measured root mean square error between the computed and simulated geometric centers. We have developed a mechanism by which the accuracy of fiducial locations in a low resolution image can be improved by using the fiducial locations in a high resolution image. Because the markers are indistinguishable it is necessary to determine their correspondence across images. We have adopted and evaluated geometric algorithms for this purpose. The inherent discreteness and partial volume effect associated with digital images limit the accuracy of fiducial alignment, which in turn limits accuracy in the alignment of anatomical targets. We have conducted experiments to assess this limitation on simulated data, CT scans of a cadaver head with implanted markers, and CT and MRI scans of human volunteers with externally attached markers. Simulations involve generating arbitrarily oriented volume images of a hemispherical model of the head, randomly picking corresponding fiducial points and targets in the images, pertubing fiducial locations with uniformly distributed error, computing transformation needed to register the images using peturbed fiducials, and measuring error in target alignment in registered images. We have compared several transformation algorithms: rigid body with (3 point) Euler angle fit, non-rigid affine transformation with n point least squares fit, rigid body with n point centroid alignment followed by Euler angle fit, and rigid body with n point least squares fit using two different approaches, namely, iterative projections and singular value decomposition.