TotalSpineSeg: automated whole-spine vertebral segmentation with high robustness to labeling errors.
Authors
Affiliations (17)
Affiliations (17)
- Department of Urology, Shaare Zedek Medical Center, Faculty of Medicine, Hebrew University, Jerusalem, Israel.
- Multiple Sclerosis Center, Sheba Medical Center, Ramat-Gan, Israel.
- Arrow Program for Medical Research Education, Sheba Medical Center, Ramat-Gan, Israel.
- NeuroPoly Lab, Institute of Biomedical Engineering, Polytechnique Montreal, Montreal, QC, Canada.
- Mila-Quebec AI Institute, Montreal, QC, Canada.
- Department of Radiology, Jefferson Integrated Magnetic Resonance Imaging Center, Thomas Jefferson University, Philadelphia, PA, USA.
- National Institute of Neurological Disorders and Stroke, National Institutes of Health (NIH), Bethesda, MD, USA.
- Department of Biomedical Engineering, University of Basel, Allschwil, Switzerland, University Hospital Basel and University of Basel, Basel, Switzerland.
- University of British Columbia (UBC), Vancouver, BC, Canada.
- Division of Neurology, Department of Medicine, St. Michael's Hospital, University of Toronto, Toronto, ON, Canada.
- Department of Neuroscience at Université de Montréal, Montréal, QC, Canada.
- Cumming School of Medicine at the University of Calgary, Calgary, AB, Canada.
- Faculty of Medicine, Tel-Aviv University, Tel-Aviv, Israel.
- NeuroPoly Lab, Institute of Biomedical Engineering, Polytechnique Montreal, Montreal, QC, Canada. [email protected].
- Mila-Quebec AI Institute, Montreal, QC, Canada. [email protected].
- Functional Neuroimaging Unit, CRIUGM, Université de Montréal, Montreal, QC, Canada. [email protected].
- Centre de recherche du CHU Sainte-Justine, Université de Montréal, Montreal, QC, Canada. [email protected].
Abstract
To develop TotalSpineSeg, a robust tool for vertebrae and intervertebral disc instance segmentation, along with spinal cord and spinal canal segmentation, capable of processing any 3D medical image regardless of contrast, resolution, or field of view, while overcoming fusion and fragmentation errors common in standard segmentation. A two-stage cascaded deep learning model was trained on 1404 MRIs. To explicitly prevent errors where adjacent vertebrae merge or a single vertebra splits, the architecture employs a novel "odd/even" representation. The second stage is conditioned on an odd-intervertebral disc mask from the first stage to assign distinct semantic classes to alternating vertebrae, enforcing accurate separation between neighboring vertebrae. Finally, a landmark-based logic uses key anatomical points (C2-C3, C7-T1, T12-L1, and L5-S) to anchor sequential labeling, while a localizer-based strategy ensures accurate identification even in truncated, short field-of-view scans. Evaluation across diverse contrasts and pathologies yielded a vertebral identification rate of 96.22% on localizer MRI and generalized to CT scans with a 92.87% identification rate. For intervertebral disc instance labeling, where vertebral separation is not mandatory, both model stages attained 99% labeling accuracy with a mean positional L2 error of 1.55 mm, surpassing existing state-of-the-art methods. TotalSpineSeg successfully generalizes to diverse 3D medical images, overcoming variations in contrast, resolution, and field of view. The proposed topological constraints effectively eliminate vertebral fusion and fragmentation errors, delivering a state-of-the-art identification rate for both vertebrae and intervertebral discs across MRI and CT modalities. Question Can a topology-constrained framework overcome heterogeneous contrasts and resolution limits, including localizers, to ensure robust, automated vertebral and disc labeling? Findings TotalSpineSeg achieved a high identification rate across diverse contrasts, localizers, and CT, surpassing state-of-the-art methods by eliminating fusion and fragmentation errors. Clinical relevance TotalSpineSeg enables fully automated biomarker extraction from clinical scans, allowing precise assessment of stenosis, lesions, and atrophy against normative ranges.