Purpose There is now convincing evidence that prostate cancer cells lack the ability to produce and accumulate citrate. Using magnetic resonance spectroscopy imaging (MRSI), regions of absent or low citrate concentration in the prostate can be visualized at a resolution of a few mm. This new advancement provides not only a tool for early diagnosis and screening but also the opportunity for preferential targeting of radiation to regions of high tumor burden in the prostate. The differences in the shape and location of the prostate between MRSI imaging and treatment have been the major obstacle in integrating MRSI in radiation therapy treatment planning. The purpose of this study is to develop a reliable method for deforming the prostate and surrounding regions from the geometry of MRSI imaging to the geometry of treatment planning, so that the regions of high tumor burden identified by the MRSI study can be faithfully transferred to the images used for treatment planning. Methods and materials Magnetic resonance spectroscopy imaging studies have been performed on 2 prostate cancer patients using a commercial MRSI system with an endorectal coil and coupling balloon. At the end of each study, we also acquired the MRI of the pelvic region at both the deformed state where the prostate is distorted by the endorectal balloon and the resting state with the endorectal balloon deflated and removed. The task is to find a three-dimensional matrix of transformation vectors for all volume elements that links the two image sets. We have implemented an optimization method to iteratively optimize the transformation vectors using a Newton-Ralphson algorithm. The objective function is based on the mutual information. The distorted images using the transformation vectors are compared with the images acquired at the resting conditions. Results and discussion The algorithm is capable of performing the registration automatically without the need for intervention. It does not require manual contouring of the organs. By applying the algorithm to multiple image sets of different patients, we found a good agreement between the images transformed from those acquired at the deformed state and those acquired at resting conditions. The computation time required for achieving the registration is in the range of a half-hour (for image size: 256 pixels × 256 pixels × 25 slices). However, the space of registration can be restricted to speed up the process. Conclusion In this article, we described a three-dimensional deformable image registration method to automatically transform images from the deformed imaging state to resting state. Our examples show that this method is feasible and useful to the treatment planning system. © 2004 Elsevier Inc.