Core Concepts
A deep learning approach called SPINEPS that can automatically segment 14 spinal structures, including vertebrae substructures, intervertebral discs, spinal cord, and spinal canal, in T2-weighted MRI scans of the whole spine.
Abstract
The authors present SPINEPS, a two-phase deep learning approach for semantic and instance segmentation of 14 spinal structures in T2-weighted whole-body MRI scans.
In the first phase, a semantic segmentation model is used to segment the scans into 14 different spinal structures, including 10 vertebral substructures, intervertebral discs, spinal cord, and spinal canal. In the second phase, a sliding window instance segmentation model is applied to the semantic segmentation to identify individual vertebrae instances.
The approach was trained and evaluated on data from the public SPIDER dataset, a subset of the German National Cohort (NAKO), and an in-house dataset. On the SPIDER test set, SPINEPS outperformed a baseline nnUNet model across various metrics, including Dice similarity coefficient, average symmetric surface distance, and instance-wise segmentation quality.
When trained only on automated annotations derived from a combination of existing segmentation models and MR-to-CT translation, the approach achieved Dice scores of 0.90 for vertebrae, 0.96 for intervertebral discs, and 0.95 for the spinal canal on a manually corrected NAKO test set. Further incorporating the manually annotated SPIDER dataset improved these scores to 0.92, 0.97, and 0.96, respectively.
Qualitative evaluation on the in-house dataset demonstrated the robustness of the approach to out-of-distribution samples from different scanners, field strengths, and spatial resolutions, as well as cases with pathologies.
Stats
The vertebra corpus Dice similarity coefficient was 0.96.
The intervertebral disc Dice similarity coefficient was 0.97.
The spinal canal Dice similarity coefficient was 0.96.
The spinal cord Dice similarity coefficient was 0.97.
Quotes
"The proposed segmentation approach offers robust segmentation of 14 spinal structures in T2w sagittal images, including the spinal cord, spinal canal, intervertebral discs, endplate, sacrum, and vertebrae."
"Training on auto-generated annotations and evaluating on manually corrected test data from the GNC yielded global dice scores of 0.900 for vertebrae, 0.960 for intervertebral discs, and 0.947 for the spinal canal."