Anatomical Localization of Intracranial Electrodes Using Synthetic MRI Generated from Computed Tomography.
Authors
Affiliations (6)
Affiliations (6)
- Department of Neurosurgery, Harvard Medical School, Mass General Brigham, Boston, Massachusetts, USA.
- Center for Brain Circuit Therapeutics, Harvard Medical School, Brigham and Women's Hospital, Boston, Massachusetts, USA.
- Athinoula A. Martinos Center for Biomedical Imaging, Harvard Medical School, Massachusetts General Hospital, Boston, Massachusetts, USA.
- Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, Massachusetts, USA.
- Department of Neurology with Experimental Neurology, Charité - Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt- Universität zu Berlin, Berlin, Germany.
- Department of Stereotactic and Functional Neurosurgery, Institute for Network Stimulation, University Hospital Cologne, Cologne, Germany.
Abstract
MRI and computed tomography (CT) are commonly combined to localize intracranial electrodes in deep brain stimulation (DBS). However, when MRI access is limited or unsafe, only CT may be available, which lacks the soft tissue contrast needed for accurate anatomical reference. SynthSR, a deep learning tool that generates synthetic T1-weighted MRI from CT, may address this limitation. We evaluated the accuracy of SynthSR-based localization compared to standard MRI-CT fusion. Preoperative CT, preoperative MRI, and postoperative CT scans were analyzed in 44 patients undergoing subthalamic nucleus (STN) DBS for Parkinson's disease (mean age = 60.7 ± 6.9 years; mean disease duration = 10.4 ± 3.9 years). SynthSR generated synthetic T1 scans from preoperative CTs. Post-operative CT was co-registered to synthetic and real MRIs, and electrodes were localized and warped to Montreal Neurological Institute space using either real or synthetic MRI as reference. SynthSR-based localizations were compared to real T1 and T1 + T2 "gold-standard" localizations using 3-dimensional Euclidean and radial distances, as well as anatomical mapping within STN subdivisions. In Montreal Neurological Institute space, the mean radial error between SynthSR- and T1 + T2-based localizations was 1.21 ± 0.77 mm, with a Euclidean distance of 1.67 ± 0.93 mm-comparable to electrode diameter. Roughly 10% of radial errors were >2 mm and 1% were >4 mm. Around 30% of Euclidean distances were >2 mm, and 3% were >4 mm. Anatomical localization concordance was 95.5% for overall STN targeting and 81.8% within STN subdivisions. DBS electrode localization using synthetic MRI generated from CT via SynthSR demonstrated preliminary feasibility for DBS electrode localization, although localization differences varied across cases. This CT-only approach may enable postoperative evaluation when MRI is unavailable or contraindicated, supporting broader access to imaging-based DBS analysis as synthetic image generation continues to improve.