J. Neurosurg. / Volume 109 / December 2008 J Neurosurg (Suppl) 109:000–000, 2008 21 I mage registration of anatomical and functional images from either the same imaging modality or different imaging modalities is performed routinely to delin- eate target volumes for radiation therapy treatment plan- ning. 2–4 The procedure is also important if one needs to evaluate the treatment outcome by comparing the images used for treatment and those taken for posttreatment eval- uation. In particular, we were interested in quantitatively monitoring the local control of the tumors treated with stereotactic GKS. 9 In that study we attempted to identify the geographical location of local control failure in corre- lation to the radiation dose distribution used for the treat- ment. Such analyses required accurate coregistration of the treatment planning image set and the posttreatment image set. Because the coregistration is usually accom- plished using an automatic registration (or image fusion) tool, the accuracy of the tool must be known to draw any conclusion on the geographical miss of the radiation delivery. In the current study we evaluated, using a phan- tom, the accuracy of the automatic coregistration tool implemented in the LGP treatment planning software. Methods We tested the image registration tool in the LGP Ver- sions 4C (Release 5.34) and 8.0 (Elekta). The version 4C needs additional modules for image registration, which are independently provided as MultiView and Image- Merge modules. Version 8.0 already includes the Image- Merge-based image registration tool. In the LGP soft- ware, a NMI algorithm is used for image coregistration. 6 The information-based coregistration algorithm tries to J. Neurosurg. / Volume 109 / December 2008 J Neurosurg 109:21–24, 2008 Image registration accuracy of GammaPlan: a phantom study Y oichi W atanabe, Ph.D., 1 anD eunYoung han, Ph.D. 2 1 Department of Therapeutic Radiology, University of Minnesota, Minneapolis, Minnesota; and 2 Radiation Oncology Department, Berkshire Medical Center, Pittsfeld, Massachusetts Object. The authors evaluated the accuracy of the automatic image coregistration function implemented in the Leksell GammaPlan treatment planning software (Version 4C with MultiView Extension and Version 8.0). Methods. The authors used a phantom with 9 landmarks (tips of thin cylindrical acrylic rods) evenly distributed in the treatment space. Two sets of images of the phantom were taken with both CT and MR imaging systems. The frst image was obtained with the phantom aligned with the scanner’s axis and the second scan was made by inten- tionally shifting and rotating the phantom relative to the scanner’s axis. The authors attempted image registration of 2 CT image sets, CT and MR image sets, and 2 MR image sets. The accuracy of image registration was evaluated by measuring the x, y, and z coordinate values of the landmarks on each image set after 2 image sets were coregistered. The authors calculated the differences of the x, y, and z values and the distance, d, between corresponding landmarks in 2 image sets. To minimize interobserver dependence of coordinate measurements, 2 physicists did measurements independently. Results. The distances, d, averaged over the 9 landmarks, were 2.63 ± 1.64 and 0.95 ± 0.25 mm for CT–CT and MR–MR image registrations, respectively. When the CT images of the air-flled phantom and MR images were coregistered, however, the algorithm performed poorly: d = 13.8 ± 1.23 mm. To remedy this, the authors undertook a 2-step process by frst performing landmark-based registration of the 2 image sets and subsequently applying the automatic registration. With this approach, the mean distance drastically improved: d = 0.74 ± 0.31 mm. When the water-flled phantom was used for CT scans, the registration accuracy of CT and MR image sets was acceptable without the 2-step registration process: d = 1.18 ± 0.36 mm. Conclusions. The accuracy of automatic registration of image sets from the same modality was within the voxel size of the scanned images. The accuracy of CT–MR image registration strongly depended on whether the phantom for CT scans was flled with air or water. This indicates the signifcant effect of the amount of common data available for a mutual information-based algorithm on the accuracy. (DOI: 10.3171/JNS/2008/109/12/S5) KeY WorDs • computed tomography • image coregistration • image fusion • magnetic resonance imaging • normalized mutual information • phantom • radiosurgery 21 Abbreviations used in this paper: GK = Gamma Knife; GKS = GK surgery; LGP = Leksell GammaPlan; NMI = normalized mutual information.