In this paper we propose an atlas-based segmentation technique for subcortical structures in 3D MR images using non-rigid image registration. Further we evaluate two separate transformation models used in non-rigid registration method, namely, Thin Plate Splines (TPS) and Cubic B Splines (CBS). The optimization technique used for the registration process was Powell's method and the similarity measure used for TPS based registration was normalized mutual information whereas normalized cross correlation was used in CBS based registration algorithm. The results of automatically segmented structures (which include ventricles, caudate nucleus and putamen) obtained via atlas-subject registration were assessed against manual segmentation, using sensitivity (S), positive predictive value (P) and Dice coefficient (D) metrics. The mean ± std values of S, P and D are 0.92 ± 0.01, 0.93 ± 0.01, 0.93 ± 0.01 respectively in case of CBS whereas 0.85 ± 0.01, 0.85 ± 0.01, 0.84 ± 0.02 are the mean ± std values of S, P and D respectively in case of TPS. Thus results indicate that the better approach to segment the subcortical structures, both in terms of speed and accuracy, is by using CBS based non-rigid registration algorithm.