Comprehensive Summary
This study formulated a publicly available Cervical Spine MRI Segmentation Dataset (CSpineSeg) and developed a deep learning segmentation model for cervical spine segmentation tasks. CSpineSeg was made by collecting 1,255 cervical spine magnetic resonance imaging (MRI) examinations from 1,232 patients (481 of which included expert-verified manual semantic segmentations). A deep learning model was then trained on the manually annotated data, and used to generate segmentations on the remaining unannotated images. The developed segmentation model achieved a Dice Coefficient of 0.929 for vertebral bodies, 0.904 for intervertebral discs, and a macro-average of 0.916. It is important to note that the segmentations used in the study were not classified or labeled by specific vertebral levels, and there was no cross referencing to verify the manual segmentation process.
Outcomes and Implications
Current manual methods of cervical spine segmentation are extremely tedious and time consuming. Deep learning alternatives have the potential to fast track this process and the diagnosis of cervical spine diseases such as degenerative spondylosis, spinal infection, and spinal tumors. In addition, there are no publicly available datasets on cervical spine MRI’s with comprehensive vertebral body and intervertebral discs segmentation to develop algorithms like this. The successful development and implementation of both the algorithm and the CSpineSeg dataset would increase clinical efficiency and allow the development of more algorithms to further this cause. Although the authors do not explicitly comment on clinical implementation, they suggest future work includes the addition of vertebral-level dataset labels, and multi-rater, cross-validated manual segmentation.