Introduction: In deep brain stimulation (DBS), precise electrode implantation is key to a successful surgical outcome. Aiming at a satisfying implantation, the location of a stimulation target in a patient's brain is usually estimated during surgical planning using pre-operative magnetic resonance images (MRI). Since targeted basal ganglia (BG) (e.g. subthalamic nucleus/STN) are hardly or partially visible on T1/T2-weighted (T1/T2w) MRI, indirect methods are used to complement MRI-based targeting, such as atlas-based registration. In our radiological department, MRI are acquired at 1.5T, and a 3D histological basal ganglia atlas (Yelnik, 2007), comprising a T1w MRI from a post-mortem specimen and its associated BG meshes, is aligned to each patient T1w MRI through piecewise-linear (PL) transforms (Bardinet, 2009). Although results are robust, nonlinear deformations could improve targeting accuracy, especially when the atlas and the patient anatomies present large morphological differences. However, nonlinear deformations (e.g. ANTS - Avants, 2008) may correctly warp high-contrast brain structures, but deform low-contrast ones (e.g. STN) in an anatomically spurious way. Thus, this study assesses an iconic-geometric nonlinear registration method to warp the 3D histological atlas into patient data, aiming at a more anatomically-robust and precise localization of the STN in DBS for Parkinson's disease. Methods: The adopted algorithm finds a simultaneous registration from the source atlas MRI and a subset of its BG meshes (e.g. thalamus, putamen, caudate nucleus) to the target patient MRI and its respective meshes, obtained from the recon-all tool of Freesurfer. This compound iconic and geometric registration is based on a unified diffeomorphic framework built upon Durrleman (2013,2014). The sought deformation is found by minimizing an objective function consisting of a sum of similarity measures between source and target data, and a regularization factor. Iconic and geometric data are interpolated by a gaussian kernel applied to points distributed across the data domain, controlling how deformations propagate in high/low contrast regions. The geometric constraints (atlas and patient meshes) ensure precise mappings of the respective brain structures and influence the deformations of surrounding structures. Thus, accurate registration of visible structures around the STN should improve its location. Results: The algorithm was tested on a cohort of 30 parkinsonian patients, whose available data included 1.5T T1/T2w MRI, and neurologist annotations based on per-operative electrophysiological recordings of STN activity, that included levels of neuronal activity and anatomical labels (e.g. STN). Atlas meshes were warped into patient space using ANTS, the PL method, and the iconic-geometric nonlinear registration. Warped STN meshes from each method were quantitatively evaluated and compared. Shape indices (size, shape fractional anisotropy, elongation) showed that the proposed approach preserves the STN anatomical properties across the population, unlike ANTS. Besides, its STN meshes present lower average intensities computed from their overlap with the T2w MRI, where the STN partially appears as a hypointensity, indicating a possibly better STN localization. Finally, the PL and iconic-geometric methods were tested against the coherence between their warped STN meshes and per-operative STN electrophysiological labels using confusion matrices. The new method proved to be more specific, presenting less false positives (points with absence of STN activity inside warped STN meshes), whereas its sensibility (related to true positives, points within the STN meshes recognized as the STN by the electrophysiology) remained competitive with the currently adopted method. Conclusions: The proposed nonlinear deformation model is suitable for the given atlas-to-patient registration task, since it improved STN location estimation precision and preserved its anatomical features.