Automated robust segmentation of the spinal canal on MRI

Background and objective Accurate segmentation of the spinal canal is important for quantitative analysis in various spinal pathologies. However, manual delineation is time-consuming and subject to inter- and intra-rater variability, especially in large multi-center studies. This work aimed to develop an automatic model to segment the spinal canal, defined as the dural sac, from T2-weighted (T2w) magnetic resonance imaging sequences. Methods This retrospective multi-site model-development study used a 3D convolutional neural network (nnUNetv2 framework) and was trained using an active learning strategy on a multi-site dataset (n > 18) with varying imaging parameters (resolution, field-of-view). The dataset included T2w images (TSE and 3D CISS) covering the cervical, thoracic, and lumbar spine from both healthy participants and patients with conditions like degenerative cervical myelopathy and traumatic spinal cord injury, including post-operative. The model was tested on unseen T2w images from two sites, including a range of acquisition protocols and clinical presentations, from mild to severe canal compression, to evaluate the performance and robustness. Comparisons were made against state-of-the-art (SOTA) models. Ethics approvals were obtained at each contributing institution; approval identifiers are reported in the main manuscript Results The proposed model achieved strong and consistent quantitative performance (e.g. mean Dice similarity coefficient of 0.95 and mean Hausdorff distance of 1.17 mm² on a healthy thoraco-lumbar test cohort), outperforming the SOTA. Notably, it showed superior robustness in challenging cases of severe spinal canal stenosis, achieving Dice scores > 0.9 where SOTA models dropped below 0.5. Conclusions The model generalized well across different MRI vendors, sequences, resolutions, and fields-of-view, for both healthy individuals and patients. The model is open-source and available in the Spinal Cord Toolbox v6.4 and higher.