those that would normally be encountered, in order to test
the algorithms at large angles.
The film was marked with three reference points which
remained in the same positions on the film throughout the
image capture with the CID camera. To test system
performance, four anatomic features were chosen on the first
image as it was displayed on the monitor. The upper incisor
is an example of the type of feature selected. In addition,
one or more specific points on each feature were selected
and stored using a cursor also shown on the screen. For
example, the edge of the upper incisor was one such point.
A region around the entire tooth was selected using a
moveable displayed window, and the digital "patch" stored on
disk. The three subsequent rotated images were displayed on
the monitor, and the same four points and patches chosen as
closely as possible, given the changes in images appearance
due to the rotation. After all feature points and image
patches were chosen, the boundary detection software
processed all the sub-images.
The main goal of this testing procedure was to determine the
accuracy with which feature boundaries could be extracted
and fitted together in the presences of rotation and
translation. However, because each anatomic point was
selected independently on each of the four orientations of
the x-ray, pointing error was a component of the final
misclosure in point coordinates. In order to separate the
pointing error from the transformation error, the following
procedure was developed.
On the first film, the x and y location of each point, for
example the incisor edge, was taken to be correct. Then,
the first film was transformed onto the second film, using a
four parameter transformation on the three fixed reference
points. The coordinates of the anatomic feature point (i.e.
incisor edge) were transformed to the coordinate system of
the second film, and inserted as an artificial point into
the digital representation of the second film. This process
was repeated for all points on the second, third, and fourth
films.
Then, each feature boundary was fit to its boundary on the
first film as determined by the boundary detection
algorithm. For example, the boundary for the upper incisor
of the second film was transformed to the corresponding
boundary on the first film, and the coordinates of the
anatomic point selected on the second film (i.e. incisor
edge) were then reported in the coordinate system of the
first film. In addition, the artificial point on the second
film was transformed, using the same parameters, to the
coordinate system of the first.
If the fit on feature boundaries were exact, the coordinates
of the artificial point would be the same as on the first
film, and the anatomic point would differ only because of
pointing error. However, because the fits were not exact
due to boundary changes under rotation, and algorithmic
- 169 -