The subthalamic nucleus (STN) within the sub-cortical region of the Basal ganglia is a crucial targeting structure for Parkinson's Deep brain stimulation (DBS) surgery. Volumetric segmentation of such small and complex structure, which is elusive in clinical MRI protocols, is thereby a pre-requisite process for reliable DBS direct targeting. While direct visualization of the STN is facilitated with advanced ultrahigh-field MR imaging (7 Tesla), such high fields are not always clinically available. In this paper, we aim at the automatic prediction of the STN region on clinical low-field MRI, exploiting dependencies between the STN and its adjacent structures, learned from ultrahigh-field MRI. We present a framework based on a statistical shape model to learn such shape relationship on high quality MR data sets. This allows for an accurate prediction and visualization of the STN structure, given detectable predictors on the low-field MRI. Experimental results on Parkinson's patients demonstrate that the proposed approach enables accurate estimation of the STN on clinical 1.5T MRI.