Precise Electrode Co-alignment in Deep Brain Stimulation Fusing Neuroimaging and Electrophysiology

Precise Electrode Co-alignment in Deep Brain Stimulation Fusing Neuroimaging and Electrophysiology Igor Varga1,2,Daniel Novak1, Dusan Urgosik3, Jan Kybic1, Filip Ruzicka3, Pavel Filip1,3,4, Robert Jech3,Andreas Horn5,6,7, Eduard Bakstein1,8 1Faculty of Electrical Engineering, Czech Technical University, Department of Cybernetics, Karlovo namest´ı 13, 121 35 Prague 2, Czech Republic 2Czech Centre for Phenogenomics, BIOCEV – Institute of Molecular Genetics, Prumyslova 595, 252 50 Vestec, Czech Republic 3Department of Neurology, First Faculty of Medicine, Charles University and General University Hospital in Prague, Katerinska 30, 120 00 Prague, Czech Republic 4Center for Magnetic Resonance Research (CMRR), University of Minnesota, Minneapolis, MN, USA 5Institute for Network Stimulation, Clinic for Stereotaxy and Functional Neurosurgery, University Hospital Cologne, Germany 6Center for Brain Circuit Therapeutics Department of Neurology, Brigham & Women’s Hospital, Harvard Medical School, Boston MA 02115, USA 7MGH Neurosurgery & Center for Neurotechnology and Neurorecovery (CNTR) at MGH Neurology Massachusetts General Hospital, Harvard Medical School, Boston, MA 02114, USA 8National Institute of Mental Health, Topolova 748, 250 67 Klecany, Czech Republic E-mail: [email protected];[email protected] Abstract. Objective: To improve the precision of electrode placement in deep brain stimulation (DBS) by creating a multimodal framework that combines neuroimaging with electrophysiological data, enabling accurate electrode co-alignment. Approach: We implemented a deep learning-based workflow that combines preoperative magnetic resonance imaging (MRI) and intraoperative microelectrode recordings (MER) for DBS electrode localisation. The workflow includes automated subthalamic nucleus (STN) segmentation using a two-step convolutional neural network (CNN), MER signal classification via a transformer encoder, and spatial coalignment through a discrete optimisation framework. The entire pipeline is integrated within a 3D Slicer plugin for real-time visualisation and analysis. Main Results: The proposed method improved electrode localisation accuracy by 0.3 mm, demonstrating improved alignment between electrophysiological and anatomical targets. Real-time visualisation facilitated interactive adjustments, while automated segmentation achieved high Dice similarity scores of 0.62 ±0.10 for the STN, sufficient for manual refinement. Significance: This multimodal approach reduces electrode placement errors, incorporates both preoperative and intraoperative data, and provides clinicians with a robust real-time tool to improve DBS results. The integration of neuroimaging and electrophysiology addresses long-standing challenges in DBS, offering a significant step toward personalised and precise neurosurgical interventions. Keywords: Deep Brain Stimulation, Precise Electrode Localisation, Neuroimaging, Electrophysiology, Subthalamic Nucleus, Optimisation Framework. Precise Electrode Co-alignment in Deep Brain Stimulation 2 1. Introduction Integrating information obtained from neuroimaging data with the one obtained from invasive electrophysiological recordings is a promising pathway to optimise electrode placement in deep brain stimulation (DBS), particularly to treat neurological disorders such as Parkinson’s disease (PD) [1]. Accurate localisation of the subthalamic nucleus (STN) - a primary target of PD DBS - during surgery is important to maximise clinical outcomes. However, its small size, deep location, and individual anatomical variability pose significant challenges for precise identification and segmentation [2, 3]. Typically, the location of the STN is first visualised using preoperative magnetic resonance imaging (MRI) or computed tomography (CT), which is used for surgical planning, that is, to determine the trajectory of the DBS electrode. While indispensable, neuroimaging techniques have limitations in spatial resolution and may introduce distortions, particularly in deep brain structures such as the STN [4]. Accurate localisation of the STN is further complicated by its low contrast from Subtantia Nigra, even if specialised sequences such as T2 or T2* are used [5]. The actual position of the STN during subsequent DBS implantation may deviate from the surgical plan due to several factors, namely stereotactic and co-registration inaccuracies, air entering the skull due to cerebrospinal fluid leakage (pneumocephalus) or postural changes. These aspects may lead to brain shift or non-linear brain deformations, complicating the accurate implantation [6, 7]. For these reasons, many centers use intraoperative invasive microelectrode recordings (MER), which can give clues about inaccuracies of the location provided previously by imaging and deviations from surgical plan and provide a more accurate STN delineation before determining the final stimulation contact positions [8]. MER has long been instrumental in confirming or refining the optimal target site for electrodes that are entered during surgery. Although the procedure can prolong operative time and carries risks such as intracranial haemorrhage and postoperative infections [9, 10, 11], MER continues to offer valuable real-time electrophysiological feedback. Typically, multiple electrodes are entered to probe a larger field of anatomical tissue, and these approaches have shown promise in improving motor outcomes in Parkinson’s disease, especially for symptoms such as tremor and rigidity [12]. Despite its promise and success, MER alone delivers clues that need to be manually integrated by experts that have a strong anatomical understanding and electrophysiological expertise, which typically requires extensive training in the operating room (OR) and depends on signal quality [13]. Together, these challenges – issues in the imaging accuracy when used alone and complicated nature of understanding MER signals – highlight the need for advanced multimodal approaches that integrate electrophysiological information obtained during MER obtained in the OR together with the neuroimaging data collected preoperatively. Discrepancies observed while merging the two sources of information can provide clues about brain shift and valuable insights when determining final implantation position of Precise Electrode Co-alignment in Deep Brain Stimulation 3 the permanent lead. Indeed, such discrepancies have shown to arise: Studies report that MRI can miss vital anatomical details of the STN in up to 20% of cases [14], and different field strengths can shift the apparent boundaries of the STN [15], notably leaving the dorsal margin inconsistently defined. Although high-field magnetic resonance imaging (7T) can reduce lateral border ambiguities, the dorsal boundary remains challenging to precisely pinpoint. In this context, MER still offers critical supplementary information, and optimising its integration with imaging is an active area of research aiming to minimise brain shift effects and other surgical uncertainties. A recent key advancement in this direction is the Lead-OR platform [16], which spatially visualises MER signals in the stereotactic space defined by MRI (and nuclei segmentations obtained from it) in real time to improve DBS targeting. Here, we build upon and significantly extend this platform by proposing an end-to-end pipeline capable of automatically enhancing electrode localisation by integrating deep learningbased STN segmentations with MER co-alignments. Convolutional neural networks (CNN) have shown strong performance in segmenting subcortical structures, including STN [17, 18]. Our approach employs a two-step CNN model for precise anatomical segmentation, followed by an optimisation framework that aligns MER signals within the segmented STN. Implemented as a 3D Slicer[19] plugin and integrated into LeadOR, this pipeline provides a robust framework for clinicians and researchers to visualise and analyse neuroimaging and electrophysiological data simultaneously. By reducing target error and accounting for intraoperative brain shift, our method aims to improve electrode targeting and ultimately improve patient outcomes in DBS. 2. Methods Our aim is to provide surgeons with real-time information about the electrode’s position relative to surrounding anatomical structures during surgery. Ideally, this process should achieve maximum precision and operate in real time. For the purpose of this article, we refer to this process as ‘electrode location’, which should not be confused with similar terms used postoperatively, i.e. in the process of localising electrodes already implanted based on postoperative imaging. In summary, to accurately localise electrodes in real time during DBS surgery, we implemented a shift estimation method that combines deep learning with classical optimisation techniques. This method integrates neuroimaging and electrophysiological data within a structured pipeline, implemented as a 3D Slicer extension. The workflow, outlined in Figure 1, can be broken down into the following components: 1. STN Segmentation from preoperative MRI: We developed a two-step convolutional neural network (CNN) model capable of automatically localising and segmenting the subthalamic nucleus (STN) from preoperative MRI scans. This approach generates a 3D surface mesh representation of the STN in stereotactic space. This automated segmentation can be manually refined using published tools available in the Lead-DBS framework [20]. Precise Electrode Co-alignment in Deep Brain Stimulation 4 2. Classification of MER signals: To address the second requirement, we developed a transformer-based encoder model that processes microelectrode recording (MER) signals, extracting electrophysiological features essential for STN localisation and classifying recording sites as either within or outside the STN. 3. Co-alignment through discrete optimisation: To integrate these two sources of information, the segmented STN shapes and classified MER data were combined within an optimisation framework designed to refine electrode placement, thereby compensating for potential brain displacement, such as intraoperative brain shift. CNN Segmentation MER Classification Preoperative MRI Manual Refinement Using WarpDrive Co-alignment framework MRI pipelineElectrophysiology pipeline Segmented STN Intraoperative MER Classified MER STN non-STN STN non-STN STN Refined STN segmentation Refined STN position Figure 1: Overview of the DBS Electrode Co-alignment Framework This framework combines preoperative MRI with intraoperative MER data, providing enhanced realtime information about electrode positions at any moment during surgery. Top row: The subthalamic nucleus (STN) is segmented from pre-operative MRI data, with optional manual refinements applied to enhance anatomical accuracy. Bottom row: Intraoperative MER signals are automatically classified into inside/outside STN groups. The co-alignment framework then aligns MER-based classifications and locaction with the refined STN segmentation. 2.1. Data and labelling Five data subsets from three data sources were used to train and validate the different steps of the method, summarised in Table 1. 2.1.1. OASIS-3 MRI Dataset We used the publicly available OASIS-3 dataset [21] to train and evaluate our segmentation model. OASIS-3 includes high-resolution structural Precise Electrode Co-alignment in Deep Brain Stimulation 5 magnetic resonance imaging, along with cognitive and clinical evaluations from 1,378 participants, which include individuals who are cognitively healthy, have mild cognitive impairment (MCI) or are diagnosed with Alzheimer’s disease (AD). For this study, we focussed on T1 and T2-weighted MRI scans acquired using 1.5T and 3T MRI scanners. The dataset includes individuals aged 42 to 95 years. We trained our CNN model using segmentation labels generated by P-Brain [17]. For validation, the STN in the test set was manually segmented using ITK-SNAP [22]. 2.1.2. External Validation: IXI Dataset For external validation of our MRI segmentation model, we used a manually labelled subset of the IXI dataset, which was previously used to develop the DISTAL atlas [23]. This data set includes T1and T2-weighted MRI scans from individuals without known health conditions, acquired using 1.5T and 3T MRI scanners. A cohort of 22 individuals (14 women, 8 men) aged 55 to 70 years (mean 63.8 ± 4.3 years) was selected to align with clinical populations seen in DBS. The STNs were manually segmented, primarily using T2-weighted images for better accuracy. These manual labels, including those for adjacent structures, were supplied by the original authors. A detailed description of the labeling procedure is available in [23]. 2.1.3. PRAGUE-MER Dataset The PRAGUE-MER dataset was used to train and evaluate the MER classification model. It contains MER recordings collected from 112 Parkinson’s disease patients undergoing DBS surgery. MER recordings were performed using one to five tungsten microelectrodes in a cross configuration, with a 2mm spacing between the electrodes. The electrodes advanced by 0.5mm between the recording positions. Ten-second recordings were collected at each position using the Leadpoint (Medtronic, MN) system, sampled at 24kHz and bandpass filtered between 500Hz and 5000Hz. Each recording position was manually labelled by the surgical team as STN or non-STN, forming the basis for classification. 2.1.4. Prague Multimodal Dataset (PD Patients Undergoing DBS) The Prague Multimodal Dataset consists of 18 participants with PD who underwent subthalamic nucleus deep brain stimulation (STN-DBS). This group included 8 women and 10 men, aged 47 to 70 years (mean 58.8 ±7.6 years). Among these participants, 12 had intraoperative O-arm CT (OARM) scans, but only 11 were used for co-alignment validation due to imaging issues. The dataset includes 3T T2-weighted MRI, 1.5T T1-weighted MRI, O-arm CT (OARM), Microelectrode recordings (MER). MER data were recorded and labelled consistently with the PRAGUE-MER dataset. Unlike PRAGUE-MER, this dataset includes both electrophysiological and imaging data, allowing for co-alignment validation. Precise Electrode Co-alignment in Deep Brain Stimulation 6 Table 1: Datasets used for training and testing of segmentation, microelectrode signal classification, and co-alignment. Usage Data source Size & Composition Modalities Notes Segmentation train OASIS-3 1,366 selected subjects T1, T2, pBrain automatic labels OASIS-3 contains various conditions including neurodegenerative Segmentation test OASIS-3 10 subjects manually labeled T1, T2, Manual labels Manual labels by a trained expert DISTAL (IXI subset) 22 subjects manually labeled T1, T2, Manual labels Manual labels by Ewert [23] MER train PRAGUE-MER 112 PD patients, 935 microelectrodes MER, MER labels MER signals labelled intraoperatively by a neurologist MER test PRAGUEMutlimodal 30 PD patients T1, T2, MER Subjects without STN recordings excluded; total 18 subjects, 28 sides Co-alignment test PRAGUEMultimodal 12 PD patients T1, T2, MER, OARM 1 subject excluded; 4 missing unilateral MER/MRI; 17 STNs in total 2.2. Model for Subcortical Segmentation We used a cascaded neural network architecture for segmentation, dividing the process into two stages. The first stage involves predicting the centre-of-mass coordinates of the STN, while the second stage focusses on delineating its boundaries to generate a 3D surface mesh. 2.2.1. MRI Preprocessing The preprocessing pipeline we employed follows the approach outlined in our previous work [24]. To standardise intensity values, we applied fuzzy c-means clustering combined with white matter intensity normalisation [25] to ensure intensity consistency between participants. White matter masks were generated using the deep Atropos algorithm [26]. All brain images were linearly co-registered in the Montreal Neurological Institute (MNI) space [27] for spatial alignment. This co-registration step maintained anatomical accuracy by applying the same transformations to both imaging data and corresponding labels, ensuring consistency and comparability across subjects. 2.2.2. STN Center Detection The center detection employs a CNN-based brain extraction method from ANTs [28], improving segmentation accuracy and processing efficiency. Detailed descriptions of the general preprocessing pipeline can be found in our previous publication [24]. A region of interest (ROI) for the STN was defined in the MNI space, extending 3 mm along each axis to account for inter-subject variability. Symmetry-based processing mirrored one hemisphere onto the other, allowing both STNs to be processed with a single model. Within this ROI, a CNN-based model predicts the centre coordinates of the STN Precise Electrode Co-alignment in Deep Brain Stimulation 7 (x, y, z), scaled to a range [0,1]. The network consists of convolutional blocks with ReLU activation and max-pooling layers, while a Sigmoid activation in the final layer constrains the output within the normalised range (network scheme is in the Supplementary Materials). 2.2.3. STN Shape Segmentation The next stage involved the segmentation of the STN. Following the initial estimate of the STN centre, we define a refined ROI around this estimated position. For the segmentation, we then employ a CNN model with architecture similar to that used in centre detection, consisting of convolutional blocks followed by max-pooling layers. The network was trained to predict the weights of the principal component analysis (PCA) representing the shape of the STN as a standard 3D mesh shape primitives [24]. This PCA-based approach provides a compact representation of shape variability around the STN centre, allowing efficient learning and accurate predictions. To improve segmentation accuracy, we incorporate the Warp Drive toolbox [20], which allows manual refinement of the STN shape. This allows clinicians to refine automated segmentation, ensuring a more accurate representation of the STN. The adjusted STN shape is then used in the co-alignment pipeline to improve alignment with recorded electrode positions, reducing errors in shift estimation. 2.3. Co-alignment of Microelectrode Recording within MRI-based STN In this step, the segmented 3D contour of the STN — optionally refined manually and derived from preoperative MRI — is automatically aligned with intraoperative MER data to obtain a more precise estimate of electrode positioning within and around the STN. This alignment process involves two key components: MER signal classification and spatial optimisation of electrode positions. 2.3.1. MER Signal Classification First, we developed a classification model to identify electrode positions within the STN, based on normalised root mean square (NRMS) values extracted at each recording step from the MER [29]. NRMS values, adapted from methods used in the Lead-OR platform [16], provide a quantitative measure of neuronal activity and are highly effective in characterising STN signals [30]. We implemented a transformer encoder model [31] to classify microelectrode recordings (MERs) as originating from inside or outside the STN. The transformer architecture is particularly well suited to modelling sequential dependencies, making it ideal for handling structured MER sequences acquired along each microelectrode trajectory. In contrast to methods that classify each depth independently, the transformer encoder integrates contextual information across multiple recording depths to improve classification robustness. Our formulation assumes that classification at a given depth ximay depend on signals from adjacent depths {xi−k, . . . , xi, . . . , xi+k}, without requiring causality Precise Electrode Co-alignment in Deep Brain Stimulation 8 constraints (i.e., both past and future inputs are available during classification). The model was trained using cross-entropy loss on expert-labelled MER segments. 2.3.2. Electrode-STN Co-alignment To improve the spatial consistency between intraoperative electrophysiological signals and preoperative imaging, we implemented a co-alignment framework based on spatial optimisation. The aim is to minimise discrepancies between the MRI-derived STN mesh and the classification of MER sites as within or outside the STN. The optimisation problem is formalised using a cost function C(x, s, α) that evaluates the quality of a given spatial transformation. Specifically, we estimate a shift vector s∈R3and an STN scaling factor α∈R+that improve the alignment between the MER points and the anatomical STN label. Optimisation penalises inconsistencies between MER classification (STN vs. non-STN) and their spatial inclusion within the transformed STN mesh. The optimal transformation parameters are obtained by solving: (ˆs, ˆα) = arg min s,α C(x, s, α) Let xidenote the original coordinates of the i-th MER recording, and let yi∈ {0,1} be its classification label. Let y∗ iindicate whether the transformed recording position xi−sis within the scaled mesh αM. Then the objective function is defined as: C(x, s, α) = 1 N N X i=1 δ(yi=y∗ i)·d(xi−s, αM) where: •δ(yi=y∗ i) is an indicator function that equals 1 when the classification and mesh inclusion disagree (i.e., a misalignment), and 0 otherwise. •d(·, αM) denotes the shortest Euclidean distance from the point to the surface of the scaled STN mesh. Minimising this cost enforces anatomical and electrophysiological overlap. Detailed step-by-step explanation of the criterion calculation is provided in the Supplementary Material. Powell’s optimisation method [32] was chosen for its robustness in nondifferentiable optimisation problems, suitable for handling discrete inclusion logic and geometric distance functions. This approach reduces manual adjustments during DBS planning and improves anatomical precision, accounting for individual anatomical variability and recording noise in the MER signal. 3. Results The primary outcome of our work is a platform that enables automatic processing and real-time co-alignment of MRI and MER data. We developed multiple extension Precise Electrode Co-alignment in Deep Brain Stimulation 9 modules within 3D Slicer to assist surgeons with planning and visualisation tasks. While based on the Lead-OR pipeline [16], our implementation significantly extends its functionality by introducing a pipeline capable of estimating the optimal transformation for aligning electrodes within the STN. Below, we provide validation results for the individual processing steps. 3.1. STN Shape Segmentation Results In the first step, generate a patient-specific STN shape based on imaging data. The effectiveness of our segmentation approach is shown in Table 2, with Dice similarity scores indicating the agreement between automated and manually refined segmentations Dice score of 0.589 (±0.112) for the left STN and 0.619 (±0.102) for the right STN on the OASIS test set, and 0.578 (±0.101) for the left STN and 0.609 (±0.095) for the right STN on the IXI dataset. As seen from the results, the segmentation accuracy is comparable to that of the pBrains model [17] when applied to the dataset used. To ensure accurate shape representation, manual refinement using WarpDrive was applied in the next step when needed to correct discrepancies in automated segmentation. Method Data STN left STN right pBrains 0.862 0.867 pBrains (OASIS labels) 3T 0.572(0.105) 0.583(0.107) CNN method (OASIS labels) 3T 0.589(0.112) 0.619(0.102) CNN method (Evert IXI labels) 3T 0.578(0.101) 0.609(0.095) Table 2: Comparison of segmentation accuracy (Dice coefficient) for different methods and datasets. 3.2. MER Signal Classification Results In the automated MER classification task, evaluated on the MER-test dataset, our model achieved an accuracy of 0.90, a sensitivity of 0.85, and a specificity of 0.91 in the test set compared to expert-based labelling – see results in Figure 2. The classification is closely aligned with manual STN labels, particularly in regions with high levels of spiking activity. While the overall accuracy remains high, the slightly lower sensitivity suggests that certain STN regions may be misclassified as outside, despite being within the STN according to expert judgement. Precise Electrode Co-alignment in Deep Brain Stimulation 16 [9] Schlaier JR, Habermeyer C, Warnat J, Lange M, Janzen A, Hochreiter A, et al. Discrepancies between the MRIand the electrophysiologically defined subthalamic nucleus. Acta Neurochirurgica. 2011 Dec;153(12):2307-18. Available from: https://doi.org/10.1007/ s00701-011-1081-7. [10] Tonge M, Ackermans L, Kocabicak E, van Kranen-Mastenbroek V, Kuijf M, Oosterloo M, et al. A detailed analysis of intracerebral hemorrhages in DBS surgeries. Clinical Neurology and Neurosurgery. 2015 Dec;139:183-7. Available from: https://www.sciencedirect.com/ science/article/pii/S0303846715300482. [11] Yang C, Qiu Y, Wang J, Wu Y, Hu X, Wu X. Intracranial hemorrhage risk factors of deep brain stimulation for Parkinson’s disease: a 2-year follow-up study. The Journal of International Medical Research. 2019 Dec;48(5):0300060519856747. Available from: https://www.ncbi.nlm. nih.gov/pmc/articles/PMC7251548/. [12] Temel Y, Wilbrink P, Duits A, Boon P, Tromp S, Ackermans L, et al. Single electrode and multiple electrode guided electrical stimulation of the subthalamic nucleus in advanced Parkinson’s disease. Operative Neurosurgery. 2007 Nov;61(5):346-57. Available from: https: //journals.lww.com/01787389-200711001-00027. [13] Gross RE, Krack P, Rodriguez-Oroz MC, Rezai AR, Benabid AL. Electrophysiological mapping for the implantation of deep brain stimulators for Parkinson’s disease and tremor. Movement Disorders. 2006;21(S14):S259-83. eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/mds.20960. Available from: https: //onlinelibrary.wiley.com/doi/abs/10.1002/mds.20960. [14] Lozano CS, Ranjan M, Boutet A, Xu DS, Kucharczyk W, Fasano A, et al. Imaging alone versus microelectrode recording–guided targeting of the STN in patients with Parkinson’s disease. Journal of Neurosurgery. 2018 Aug;130(6):1847-52. Publisher: American Association of Neurological Surgeons Section: Journal of Neurosurgery. Available from: https://thejns. org/view/journals/j-neurosurg/130/6/article-p1847.xml. [15] Verhagen R, Schuurman PR, van den Munckhof P, Contarino MF, de Bie RMA, Bour LJ. Comparative study of microelectrode recording-based STN location and MRI-based STN location in low to ultra-high field (7.0 T) T2-weighted MRI images. Journal of Neural Engineering. 2016 Dec;13(6):066009. Available from: http://stacks.iop.org/1741-2552/ 13/i=6/a=066009?key=crossref.4c8e8ac50a57e7025280bc00d2f9b419. [16] Oxenford S, Roediger J, Neudorfer C, Milosevic L, G¨uttler C, Spindler P, et al. Lead-OR: A multimodal platform for deep brain stimulation surgery. eLife. 2022 May;11:e72929. [17] Manj´on JV, Bert´o A, Romero JE, Lanuza E, Vivo-Hernando R, Aparici-Robles F, et al. pBrain: A novel pipeline for Parkinson related brain structure segmentation. NeuroImage: Clinical. 2020 Jan;25:102184. Available from: http://www.sciencedirect.com/science/article/ pii/S221315822030022X. [18] Baniasadi M, Petersen MV, Gon¸calves J, Horn A, Vlasov V, Hertel F, et al. DBSegment: Fast and robust segmentation of deep brain structures considering domain generalization. Human Brain Mapping. 2022 Oct;44(2):762-78. Available from: https://www.ncbi.nlm.nih.gov/pmc/ articles/PMC9842883/. [19] Fedorov A, Beichel R, Kalpathy-Cramer J, Finet J, Fillion-Robin JC, Pujol S, et al. 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magnetic Resonance Imaging. 2012 Nov;30(9):1323-41. Available from: https://linkinghub.elsevier.com/retrieve/pii/ S0730725X12001816. [20] Oxenford S, R´ıos AS, Hollunder B, Neudorfer C, Boutet A, Elias GJB, et al. WarpDrive: Improving spatial normalization using manual refinements. Medical Image Analysis. 2024 Jan;91:103041. Available from: https://www.sciencedirect.com/science/article/pii/ S1361841523003018. [21] LaMontagne PJ, Benzinger TL, Morris JC, Keefe S, Hornbeck R, Xiong C, et al.. OASIS-3: Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Precise Electrode Co-alignment in Deep Brain Stimulation 17 Disease. medRxiv; 2019. ISSN: 1901-4902 Pages: 2019.12.13.19014902. Available from: https://www.medrxiv.org/content/10.1101/2019.12.13.19014902v1. [22] Yushkevich PA, Piven J, Hazlett HC, Smith RG, Ho S, Gee JC, et al. User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability. NeuroImage. 2006 Jul;31(3):1116-28. Available from: https://www.sciencedirect.com/ science/article/pii/S1053811906000632. [23] Ewert S, Plettig P, Li N, Chakravarty MM, Collins DL, Herrington TM, et al. Toward defining deep brain stimulation targets in MNI space: A subcortical atlas based on multimodal MRI, histology and structural connectivity. NeuroImage. 2018 Apr;170:271-82. Available from: https://www.sciencedirect.com/science/article/pii/S105381191730407X. [24] Varga I, Bakstein E, Gilmore G, May J, Novak D. Statistical segmentation model for accurate electrode positioning in Parkinson’s deep brain stimulation based on clinical low-resolution image data and electrophysiology. PLOS ONE. 2024 Mar;19(3):e0298320. Publisher: Public Library of Science. Available from: https://journals.plos.org/plosone/article?id=10. 1371/journal.pone.0298320. [25] Reinacher PC, V´arkuti B, Kr¨uger MT, Piroth T, Egger K, Roelz R, et al. Automatic Segmentation of the Subthalamic Nucleus: A Viable Option to Support Planning and Visualization of PatientSpecific Targeting in Deep Brain Stimulation. Operative Neurosurgery. 2019 Nov;17(5):497-502. Available from: https://academic.oup.com/ons/article/17/5/497/5375882. [26] Avants BB, Tustison NJ, Wu J, Cook PA, Gee JC. An open source multivariate framework for n-tissue segmentation with evaluation on public data. Neuroinformatics. 2011 Dec;9(4):381-400. [27] Grabner G, Janke AL, Budge MM, Smith D, Pruessner J, Collins DL. Symmetric Atlasing and Model Based Segmentation: An Application to the Hippocampus in Older Adults. In: Hutchison D, Kanade T, Kittler J, Kleinberg JM, Mattern F, Mitchell JC, et al., editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2006. vol. 4191. Berlin, Heidelberg: Springer Berlin Heidelberg; 2006. p. 58-66. Series Title: Lecture Notes in Computer Science. [28] Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods. 2021 Feb;18(2):203-11. Publisher: Nature Publishing Group. Available from: https://www.nature.com/articles/ s41592-020-01008-z. [29] Moran A, Bar-Gad I, Bergman H, Israel Z. Real-time refinement of subthalamic nucleus targeting using Bayesian decision-making on the root mean square measure. Movement Disorders. 2006;21(9):1425-31. eprint: https://movementdisorders.onlinelibrary.wiley.com/doi/pdf/10.1002/mds.20995. Available from: https://movementdisorders.onlinelibrary.wiley.com/doi/abs/10.1002/mds. 20995. [30] Thompson JA, Oukal S, Bergman H, Ojemann S, Hebb AO, Hanrahan S, et al. Semi-automated application for estimating subthalamic nucleus boundaries and optimal target selection for deep brain stimulation implantation surgery. Journal of Neurosurgery. 2018 May;130(4):1224-33. Publisher: American Association of Neurological Surgeons Section: Journal of Neurosurgery. Available from: https://thejns.org/view/journals/j-neurosurg/130/4/article-p1224. xml. [31] Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, et al.. Attention Is All You Need. arXiv; 2023. ArXiv:1706.03762. Available from: http://arxiv.org/abs/1706.03762. [32] Powell MJD. An efficient method for finding the minimum of a function of several variables without calculating derivatives. The Computer Journal. 1964 Feb;7(2):155-62. Available from: https://academic.oup.com/comjnl/article-lookup/doi/10.1093/comjnl/7.2.155. [33] Karthick PA, Wan KR, An Qi AS, Dauwels J, King NKK. Automated detection of subthalamic nucleus in deep brain stimulation surgery for Parkinson’s disease using microelectrode recordings and wavelet packet features. Journal of Neuroscience Methods. 2020 Sep;343:108826. Available from: https://www.sciencedirect.com/science/article/pii/S0165027020302491. Precise Electrode Co-alignment in Deep Brain Stimulation 18 [34] Coelli S, Levi V, Del Vecchio Del Vecchio J, Mailland E, Rinaldo S, Eleopra R, et al. An intraoperative feature-based classification of microelectrode recordings to support the subthalamic nucleus functional identification during deep brain stimulation surgery. Journal of Neural Engineering. 2021 Feb;18(1):016003. Available from: https://iopscience.iop.org/article/ 10.1088/1741-2552/abcb15. [35] Lourens MAJ, Meijer HGE, Contarino MF, Van Den Munckhof P, Schuurman PR, Van Gils SA, et al. Functional neuronal activity and connectivity within the subthalamic nucleus in Parkinson’s disease. Clinical Neurophysiology. 2013 May;124(5):967-81. Available from: https://linkinghub.elsevier.com/retrieve/pii/S1388245712007158. [36] Khan MF, Mewes K, Gross RE, ˇ Skrinjar O. Assessment of Brain Shift Related to Deep Brain Stimulation Surgery. Stereotactic and Functional Neurosurgery. 2008;86(1):44-53. Available from: https://karger.com/SFN/article/doi/10.1159/000108588. [37] Wu YX, Xiang W, Wang JJ, Liu XM, Yi DY, Tian H, et al. A Modified Dura Puncture Procedure to Reduce Brain Shift in Deep Brain Stimulation Surgery: One Institution’s Experience. Frontiers in Neurology. 2022 Feb;13:845926. Available from: https://www.frontiersin.org/ articles/10.3389/fneur.2022.845926/full. [38] Petersen EA, Holl EM, Martinez-Torres I, Foltynie T, Limousin P, Hariz MI, et al. Minimizing Brain Shift in Stereotactic Functional Neurosurgery. Operative Neurosurgery. 2010 Sep;67(3):ons213-21. Available from: https://journals.lww.com/ 01787389-201009001-00028. [39] Luj´an JL, Noecker AM, Butson CR, Cooper SE, Walter BL, Vitek JL, et al. Automated 3Dimensional Brain Atlas Fitting to Microelectrode Recordings from Deep Brain Stimulation Surgeries. Stereotactic and Functional Neurosurgery. 2009;87(4):229-40. Available from: https: //karger.com/SFN/article/doi/10.1159/000225976.