Quantitative evaluation of a cone‐beam computed tomography–planning computed tomography deformable image registration method for adaptive radiation therapy

Deformable (non‐rigid) registration is an essential tool in both adaptive radiation therapy and image‐guided radiation therapy to account for soft‐tissue changes during the course of treatment. The evaluation method most commonly used to assess the accuracy of deformable image registration is qualitative human evaluation. Here, we propose a method for systematically measuring the accuracy of an algorithm in recovering artificially introduced deformations in cases of rigid geometry, and we use that method to quantify the ability of a modified basis spline (B‐Spline) registration algorithm to recover artificially introduced deformations. The evaluation method is entirely computer‐driven and eliminates biased interpretation associated with human evaluation; it can be applied to any chosen method of image registration. Our method involves using planning computed tomography (PCT) images acquired with a conventional CT simulator and cone‐beam computed tomography (CBCT) images acquired daily by a linear accelerator–mounted kilovoltage image system in the treatment delivery room. The deformation that occurs between the PCT and daily CBCT images is obtained using a modified version of the B‐Spline deformable model designed to overcome the low soft‐tissue contrast and the artifacts and distortions observed in CBCT images. Clinical CBCT images and contours of phantom and central nervous system cases were deformed (warped) with known random deformations. In registering the deformed with the non‐deformed image sets, we tracked the algorithm's ability to recover the original, non‐deformed set. Registration error was measured as the mean and maximum difference between the original and the registered surface contours from outlined structures. Using this approach, two sets of tests can be devised. To measure the residual error related to the optimizer's convergence performance, the warped CBCT image is registered to the unwarped version of itself, eliminating unknown factors such as noise and positioning errors. To study additional errors introduced by artifacts and noise in the CBCT image, the warped CBCT image is registered to the original PCT image. Using a B‐Spline deformable image registration algorithm, mean residual error introduced by the algorithm's performance on noise‐free images was less than 1 mm, with a maximum of 2 mm. The chosen deformable image registration model was capable of accommodating significant variability in structures over time, because the artificially introduced deformation magnitude did not significantly influence the residual error. On the second type of test, noise and artifacts reduced registration accuracy to a mean of 1.33 mm and a maximum of 4.86 mm. The accuracy of deformable image registration can be easily and consistently measured by evaluating the algorithm's ability to recover artificially introduced deformations in rigid cases in which the true solution is known a priori. The method is completely automated, applicable to any chosen registration algorithm, and does not require user interaction of any kind. PACS numbers: 87.57.Gg, 87.57.Ce, 87.62.+n