Deep Learning-Based Auto-Segmentation of Spinal Cord and Thecal Sac In Stereotactic Body Radiotherapy for Spinal Metastases
Abstract
Purpose
Stereotactic body radiotherapy (SBRT) for spinal metastases requires accurate delineation of the spinal cord and thecal sac to maximize treatment efficacy while minimizing toxicity. Deep learning-based segmentation models improve contouring efficiency and reduce variation. However, evidence specific to spinal cord and thecal sac delineation in spine SBRT remains limited. Therefore, we developed a deep learning-based automatic contouring model for the spinal cord and thecal sac and evaluated their segmentation accuracy to facilitate standardized delineation.
Methods
In this single-institution retrospective study, we analyzed 288 patients who underwent spine SBRT, using T1- and T2-weighted MRI (T1WI/T2WI) with expert-delineated 163 spinal cord contours and 125 thecal sac contours for analysis. 3D U-Net was trained under three input models: T1WI alone, T2WI alone, and combined T1WI+T2WI. The dataset was split into training, validation, and test sets in a 7:1:2 ratio. Accuracy was assessed using the Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance (HD95). For the spinal cord, a DSC ≥ 0.70 was predefined as clinically acceptable. Two-sided Wilcoxon signed-rank tests with Holm–Bonferroni correction were used (α=0.05).
Results
For the T1WI/T2WI/T1WI+T2WI models, the mean DSC (±SD) was 0.82±0.07/0.82±0.06/0.84±0.06 for the spinal cord and 0.84±0.07/0.85±0.06/0.87±0.05 for the thecal sac, respectively. In the best-performing model, all test cases achieved DSC ≥ 0.70 for the spinal cord and DSC ≥ 0.78 for the thecal sac.The mean HD95 (±SD) was 1.72±1.91/2.00±3.45/1.39±0.61 mm for the spinal cord and 2.73±1.83/2.54±1.35/2.28±1.35 mm for the thecal sac. (DSC: spinal cord, T1WI+T2WI vs T2WI, p=0.02; the thecal sac, T1WI+T2WI vs T1WI, p=0.002; HD95: the thecal sac, T1WI+T2WI vs T1WI, p=0.02.)
Conclusion
The proposed deep learning models achieved clinically acceptable spinal cord segmentation and consistently high thecal sac accuracy. Deep learning–based automatic segmentation may support spinal cord and thecal sac delineation in spine SBRT by improving efficiency and standardization.