The solution becomes unstable with the use of only ten or 
six boundary pixels. There is too much variation in shape 
and orientation to obtain a realistic fit with fewer than 20 
pixels. When all types of transformations are considered, 
the results are significantly worse for the larger rotations 
of the x-ray film. However, when only the six parameter 
transformations are considered, the results actually improve 
on the film with the highest rotation. The lower order 
transformations were not able to handle the changes in 
boundary configuration resulting from large amounts of 
rotation. 
The average misclosure vector for all of the six parameter 
fits on 20 pixels was approximately 0.5 pixels, or 
approximately 0.35 mm at the scale of the x-ray. On the 
best fit, the transformation of film 4 onto £ilm 1, the 
results averaged 0.25 pixel, or about 0.17 mm. 
5. TESTS WITH A SERIES OF X-RAY FILMS 
5.1 Data Collection 
A series of three cranial x-rays, with individual films 
spaced at one year intervals, was chosen to test the feature 
matching algorithm while including the effects of change in 
shape and size. The first film was marked with a set of 
four registration points. The second and third films were 
then registered on the first film along the, anatomic 
structure known as the anterior cranial base and the 
registration points on the first film were transferred to 
the second and third films. 
The three x-rays were imaged with the CID camera and the 
resulting digital images stored in the system. The first 
film was positioned under the camera and the locations of 
the four registration points were transferred to the light 
box surface. The subsequent x-ray films were then 
physically aligned to those registration marks before 
imaging. Therefore, all digital images were obtained in the 
same superimposed anatomic coordinate system in which the 
comparison manual digitizing would be performed. This 
anatomic superimposition did not affect the final results of 
the comparisons. it was done so that the feature 
displacements would have some anatomic validity and not 
merely be functions of the imaging orientation. 
The first digital image was displayed on the monitor and a 
series of anatomic features chosen, as described earlier, 
using the displayed window and cursor. However, the second 
and third films did not have any operator chosen features. 
The system software analyzed those images, locating the same 
series of features as on the first film, and computing 
feature displacements from image to image. 
- 171 =