Fully Automated Segmentation of Anatomical Subregions of the Thoracolumbar Spine on Computed Tomography.
Authors
Affiliations (12)
Affiliations (12)
- Machine Intelligence in Clinical Neuroscience and Microsurgical Neuroanatomy (MICN) Laboratory, Department of Neurosurgery, Clinical Neuroscience Center, University Hospital Zurich, University of Zurich, Zurich, Switzerland.
- Department of Neurosurgery, Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland.
- Department of Oncology, University of Oxford, Oxford, UK.
- Department of Neurosurgery, University Hospital Zurich, Zurich, Switzerland.
- Department of Neurosurgery, University Hospital Frankfurt, Frankfurt, Germany.
- Department of Clinical Neuroscience, Karolinska Institutet, Stockholm, Sweden.
- Capio Spine Center Stockholm, Löwenströmska Hospital, Upplands Väsby, Sweden.
- Department of Neurosurgery, University Medical Center Utrecht, Utrecht, The Netherlands.
- Machine Intelligence in Clinical Neuroscience and Microsurgical Neuroanatomy (MICN) Laboratory, Department of Neurosurgery, Clinical Neuroscience Center, University Hospital Zurich, University of Zurich, Zurich, Switzerland. [email protected].
- Department of Neurosurgery, University Hospital Zurich, Zurich, Switzerland. [email protected].
- Department of Clinical Neuroscience, Karolinska Institutet, Stockholm, Sweden. [email protected].
- Capio Spine Center Stockholm, Löwenströmska Hospital, Upplands Väsby, Sweden. [email protected].
Abstract
Segmentations of the vertebral column that include anatomical subregions can be used for patient education, pedicle screw planning, or radiomic feature extraction for spinal surgery. Deep learning has proven successful in tackling medical image segmentation; therefore, we aim to train a multiclass vertebral subregion segmentation model with different model architectures. The corpus, lamina, pedicle, superior and inferior articular process, transverse process, and spinous process of 118 thoracolumbar spine CTs were manually segmented. Then, a standard nnU-Net, a nnU-Net with residual encoder (ResEnc nnU-Net), and a shifted window U-Net transformer (SwinUNETR) were trained for semantic segmentation. All models were trained in fivefold cross-validation. A geometric approach was also used for comparison. External validation was performed on a different dataset. On an external validation dataset, the standard nnU-Net achieved the highest overlap performance across all seven vertebral subregions, with Dice scores of 0.993 ± 0.012 for the vertebral body corpus, 0.989 ± 0.023 for the pedicles, 0.991 ± 0.021 for the lamina, 0.992 ± 0.017 for the spinous process, 0.989 ± 0.025 for the transverse process, 0.985 ± 0.034 for the superior articular process, and 0.980 ± 0.046 for the inferior articular process. The Residual-Encoder nnU-Net achieved comparable but slightly lower performance (Dice 0.938-0.979 across subregions), while SwinUNETR showed a more pronounced decline (Dice 0.824-0.918), particularly in boundary-level agreement. The geometric approach of Blomenkamp et al. performed markedly worse than all deep learning-based methods, with Dice scores ranging from 0.382 for the lamina to 0.889 for the jointly segmented corpus and pedicle region.