Evaluation of magnetic resonance imaging and deep learning-based synthetic computed tomography for calcified intradural tumors - importance of domain-specific training and validation of synthetic imaging methods for clinical application.
Authors
Affiliations (7)
Affiliations (7)
- Department of Neurosurgery, H-OCH Health Ostschweiz & St. Gallen Medical School, Cantonal Hospital of St. Gallen, St. Gallen, Switzerland.
- Interdisciplinary Spine Center, H-OCH Health Ostschweiz & St. Gallen Medical School, Cantonal Hospital of St. Gallen, St. Gallen, Switzerland.
- Machine Intelligence in Clinical Neuroscience & Microsurgical Neuroanatomy (MICN) Laboratory, Department of Neurosurgery, Clinical Neuroscience Center, University Hospital Zurich, University of Zurich, Frauenklinikstrasse 10, CH-8091, Zurich, Switzerland.
- Department of Radiology and Nuclearmedicin, H-OCH Health Ostschweiz & St. Gallen Medical School, Cantonal Hospital of St. Gallen, St. Gallen, Switzerland.
- Department of Neurosurgery, H-OCH Health Ostschweiz & St. Gallen Medical School, Cantonal Hospital of St. Gallen, St. Gallen, Switzerland. [email protected].
- Interdisciplinary Spine Center, H-OCH Health Ostschweiz & St. Gallen Medical School, Cantonal Hospital of St. Gallen, St. Gallen, Switzerland. [email protected].
- Machine Intelligence in Clinical Neuroscience & Microsurgical Neuroanatomy (MICN) Laboratory, Department of Neurosurgery, Clinical Neuroscience Center, University Hospital Zurich, University of Zurich, Frauenklinikstrasse 10, CH-8091, Zurich, Switzerland. [email protected].
Abstract
For intradural spinal tumors, information on the degree of calcification is helpful to plan the surgery. Novel deep-learning algorithms allow to generate synthetic computed tomography (CT) images from magnetic resonance imaging (MRI). We conducted a prospective observational cohort study, including n = 105 patients with spinal pathologies between 07/2022 - 09/2023, to validate the accuracy of BoneMRI (MRIGuidance BV©, Utrecht, the Netherlands). Patients underwent both conventional CT and MRI; synthetic CT images were generated from MRI source data with artificial intelligence (AI). For the scope of this post-hoc analysis, only patients with intradural tumors were selected. Five patients with intradural tumors of the spine were included (mean age 67.8 years; 4 (80%) female). The tumors were visible on 5/5 conventional CT images (100%), on average 19.6 × 11.6 mm in size and 4/5 (80%) were densely calcified (mean Hounsfield units (HU) 463.6). Although well-visible on the T1w/T2w/BoneMRI source data, none of the tumors showed up (0%) on synthetic CT. Visible tumor dimensions were 0 mm in both axial (p < 0.001) and sagittal planes (p = 0.017), with an average density of 20.9 HUs (p = 0.034). BoneMRI generated synthetic CT is a promising, radiation-free alternative to conventional CT. Intradural tumors - even those with dense calcifications - were not visualized by synthetic CT images, highlighting that this novel technology is currently not able to capture lesions outside its main scope. Our analysis demonstrates powerfully that synthetic imaging must be cautiously applied to populations for which it was developed and validated, and that any extrapolation can be clinically misleading.