Points in SPECT images homologous to points in MR images may not be picked with the same anatomical accuracy. Knowledge of the nature of the information given by brain SPECT images and its poor resolution only allows the accurate selection of points with respect to large structures. It is known that there is no cerebral perfusion in the air outside the skull. The brain surface may therefore be easily outlined by thresholding the voxel values. The gross structure of the entire brain displayed in this way may then yield a surface from which unique points corresponding to the maxima, minima, and saddle points may be used to define a rigid body for registration with homologous points in the associated MR image. Similarly, the uptake ratio of gray to white matter is used to visually delineate gross structures and yield points of maximum and minimum curvature. It is in this manner that the nature of the information provided by the SPECT images limits the homology between the chosen corresponding points. The operator's eye and diagnostic experience may be thought of as another component ``black box'' in the total cascaded linear system which confounds the total imaging system's PSRF. A PSRF which includes the interaction of a human operator may also be measured. It is simply the distribution of positions which results when a person is asked to locate the mean in an image of a point source. The total imaging system in this context would include the PSRF components of the eye and the mental or interpretive faculties of an observer which include complex dependencies on luminance, contrast operator experience and other factors [Hil88]. The reduction of the dependence of the registration accuracy of the ILM technique on a physical number represented by the homology error is nevertheless made here to obtain a simple numerical descriptor of the error in point pair homology to allow the study of registration accuracy. With exact point correspondence as input, a more complete cascaded linear system which produces the quantitative FWHM measure of homology error includes a PSRF which accounts for human interaction. This may be seen in figure 2.6. The inclusion of human interaction as a component in the total imaging system may be evaluated in terms of receiver operator characteristic (ROC) analysis [Met79][Hil88].
