A quantitative approach for assessing the accuracy of the alignment process is needed. The further the selected points are from the ideal (i.e. when the sum of the squares of the residuals is zero), the further apart the points appear when plotted on the transformed image. However this does not help the user to decide when to reject the transformation due to bad fiducial point placement and try again. A simple numerical measure, independent of the number of points used, that showed the uncertainty in the transformation would probably be sufficient. One possibility is to use the sum of the squares of the residuals from the solution of the normal system divided by the number of points used. Whatever measure is adopted, work would be required to assess the normal variation in the figure and at what point to flag a bad transformation.
The use of a tablet as the input method rather than a mouse should be explored. In general people have more control over a pen than a mouse. For example preventing a cursor moving one or two pixels while clicking on a button is hard, whereas while pressing down on a tablet with a stylus this is not a problem.
The selection of fiducial points in the portal image, after histogram equalization has been applied is still difficult. The inclusion of more advanced portal image processing techniques such as contrast limited adaptive histogram equalization to improve the visibility of portal anatomy should be considered. This should make the selection of fiducial points easier, faster and more accurate.
Before using the program to correct patient alignment pro-actively, the registration transformation would need converting into machine geometry. Besides knowing the pixel size and the magnification of the two images, this also requires the coordinates of the centre of rotation of the image. One method of identifying the centre of rotation, which is also the beam centre would be to get the user to identify the four corners of the beam on the image, from which the centre could be calculated. However this method would not work on beams where the corners are obscured by shielding blocks or a multi-leaf collimator. Alternatively this could be assumed to be in the centre of the image, and the alignment of the on-line portal imaging system incorporated into the quality assurance programme.
The time required to align an image file is dependent on the experience of the user. Currently it takes about two minutes for a proficient user to align two images. This includes the loading and image processing times. If an image registration tool of this nature was to be use clinically for the pro-active correction of poor patient alignment, this would need to be reduced. Two areas that could yield significant reductions in the alignment time are a larger display and more processing power (or a reduction in current demands by increasing computational efficiency).
A larger display (i.e. screen resolution) enabling both the portal and simulator images to be displayed simultaneously in their entirety, would save the user form having to scroll the two images around looking for suitable fiducial points.
Another major overhead is the time required to do the simple image processing such as rotations/flips, histogram equalization and the transformation. Either the introduction of a workstation class machine or rewriting of the program in a compiled language such as C or C++ would reduce the time spent waiting for the computer to finish processing, reducing the registration time.
The results of the phantom studies show that the accuracy of the technique is similar to reports in the literature, though in clinical practice the accuracy of the alignment is unlikely to be this high. The size of both intra and inter-user variation in the alignment process is small enough to offer improvements in patient alignment, that show errors of more than 10mm. Studies of portal images have shown that these could account for as many as 15% of all treatments (Lam, Partowmah, Lee, Wharam & Lam 1987).
The technique appears suitable for the verification of patient position in external beam radiotherapy treatment. The small size of the variations in the alignment transformation indicate that the algorithms are robust, and the small size of the absolute errors offers the possibility of being able to improve patient setup.