Endorectal (ER) coil-based magnetic resonance imaging (MRI) and magnetic resonance spectroscopic imaging (MRSI) is often used to obtain anatomic and metabolic images of the prostate and to accurately identify and assess the intraprostatic lesions. Recent advancements in high-field (3 Tesla or above) MR techniques affords significantly enhanced signal-to-noise ratio and makes it possible to obtain high-quality MRI data. In reality, the use of rigid or inflatable endorectal probes deforms the shape of the prostate gland, and the images so obtained are not directly usable in radiation therapy planning. The purpose of this work is to apply a narrow band deformable registration model to faithfully map the acquired information from the ER-based MRI/MRSI onto treatment planning computed tomography (CT) images. A narrow band registration, which is a hybrid method combining the advantages of pixel-based and distance-based registration techniques, was used to directly register ER-based MRI/MRSI with CT. The normalized correlation between the two input images for registration was used as the metric, and the calculation was restricted to those points contained in the narrow bands around the user-delineated structures. The narrow band method is inherently efficient because of the use of a priori information of the meaningful contour data. The registration was performed in two steps. First, the two input images were grossly aligned using a rigid registration. The detailed mapping was then modeled by free form deformations based on B-spline. The limited memory Broyden-Fletcher-Goldfarb-Shanno algorithm (L-BFGS), which is known for its superior performance in dealing with high-dimensionality problems, was implemented to optimize the metric function. The convergence behavior of the algorithm was studied by self-registering an MR image with 100 randomly initiated relative positions. To evaluate the performance of the algorithm, an MR image was intentionally distorted, and an attempt was then made to register the distorted image with the original one. The ability of the algorithm to recover the original image was assessed using a checkerboard graph. The mapping of ER-based MRI onto treatment planning CT images was carried out for two clinical cases, and the performance of the registration was evaluated. A narrow band deformable image registration algorithm has been implemented for direct registration of ER-based prostate MRI/MRSI and CT studies. The convergence of the algorithm was confirmed by starting the registration experiment from more than 100 different initial conditions. It was shown that the technique can restore an MR image from intentionally introduced deformations with an accuracy of approximately 2 mm. Application of the technique to two clinical prostate MRI/CT registrations indicated that it is capable of producing clinically sensible mapping. The whole registration procedure for a complete three-dimensional study (containing 256 x 256 x 64 voxels) took less than 15 min on a standard personal computer, and the convergence was usually achieved in fewer than 100 iterations. A deformable image registration procedure suitable for mapping ER-based MRI data onto planning CT images was presented. Both hypothetical tests and patient studies have indicated that the registration is reliable and provides a valuable tool to integrate the ER-based MRI/MRSI information to guide prostate radiation therapy treatment.