FIGURE 4 Use of chamfer matching for field edge matching. (A) Exaggerated mismatch prior to optimization. The computer measures the average gray value in the distance transform under the drawing. (B) By rotating, translating, and scaling the contour and minimizing the average gray value, matching takes place. Because of the high quality of the segmented features, chamfer matching works extremely reliably and accurately for this application.
Anatomy Matching and Clinical Application
Detection of anatomical structures in electronic portal images is difficult because of the low contrast of the bony structures. Both a top-hat filter  and a multiscale medial axis filter [14,17] can be used for a low-quality segmentation of the bony anatomy (Figs 5A and 5B). However, it turns out that the chamfer matching algorithm works reasonably well even with such a poor-quality segmentation (Figs 5C and 5D). However, the likelihood that the chamfer matching algorithm is caught in a local minimum increases with a poor segmentation quality. For this reason, it is highly important that the matching result be verified visually and that the operator can manually adjust the result if required. For AP images (such as the ones shown in Figs 3 and 5) the matching reliability is over 90% . However, for other anatomy, the reliability is lower. The complete analysis procedure takes less than 1 s on a modern low-cost computer. A portal image analysis procedure based on chamfer matching has been in clinical use since 1992 and has been used for thousands of images. The analysis results of multiple beams are averaged to obtain an estimate of the patient misalignment in 3D. Using this methodology, it is possible to correct the patient setup on-line, i.e., take an image with a short exposure, analyze the image(s), and correct the patient setup. However, such a procedure is only considered cost effective for patients who are treated in a few fractions. For other patients, the image analysis is performed off-line (i.e., in the evening hours) and the results are fed into a computerized decision rule. This protocol is aimed at eliminating systematic errors in patient positioning, without influencing the day-today variations. With this protocol, the incidence of systematic patient position errors larger than 5 mm has been reduced from 30% to about 1%.
Was this article helpful?