In nonrigid image registration two problems make a challenging task. First the voxel intensity similarity and anatomical image similarity will not be equivalent. Second during imaging process some distortions such as unwanted rotations and monotonic gray-level fields may occur. Due to these problems image registration quality gets affected. To avoid these problems and to calculate efficiently Uniform Spherical Region Descriptor(USRD) method is proposed. The USRD can be calculated by comparing the gradient orientation operations and voxel-wise intensities. This method is invariant in rotation and monotonic gray level fields, so that the image can be registered without any distortions. It is an anatomical region descriptor which can encode the anatomical information around each voxel of the image. To drive the registration process, USRD feature is combined with Markov random field labelling framework in which energy function is defined for registration. The energy fuction can be maximized by using alpha expansion algorithm. The USRD features can be directly compared with other widely used features for image registration. To identify the defected region accurately normalized correlation method is used. The sample images are taken from the database Brainweb and Internet Brain Segmentation Repository respectively. Images are segmented and one of the image can be used as a template. Back propagation network is used to find the performance. © 2013 IEEE.