by Georg Hille, Sylvia Glaßer, Klaus Tönnies
Abstract:
In clinical routine, spine pathologies can be most often deduced from the vertebral body shape, position and orientation. Additionally, per-vertebra spatial information could be used in intervention planning and surgical navigation. Especially in vertebral metastasis treatment, MRI is inalienable, and therefore, segmentation methods are developed for spine MRI. Our approach starts with a simple user-assisted initialization Then intensity and edge features are combined for a subsequent hybrid level-set segmentation. We evaluated our method on highly anisotropic clinical routine spine MRI datasets, containing 34 vertebrae, both healthy and pathological. We achieved a 3D Dice coefficient of 84.8\% and a mean surface-to-surface distance of 1.29 ± 0.42mm with regard to a manually created ground truth segmentation. The main advantages of our method are precise segmentation results on clinical routine images within reasonable processing time and with minimal user interaction.
Reference:
Hybrid Level-Sets for Vertebral Body Segmentation in Clinical Spine MRI (Georg Hille, Sylvia Glaßer, Klaus Tönnies), In Procedia Computer Science, volume 90, 2016.
Bibtex Entry:
@article{hille_hybrid_2016,
title = {Hybrid {Level}-{Sets} for {Vertebral} {Body} {Segmentation} in {Clinical} {Spine} {MRI}},
volume = {90},
issn = {1877-0509},
url = {http://www.sciencedirect.com/science/article/pii/S1877050916311838},
doi = {https://doi.org/10.1016/j.procs.2016.07.005},
abstract = {In clinical routine, spine pathologies can be most often deduced from the vertebral body shape, position and orientation. Additionally, per-vertebra spatial information could be used in intervention planning and surgical navigation. Especially in vertebral metastasis treatment, MRI is inalienable, and therefore, segmentation methods are developed for spine MRI. Our approach starts with a simple user-assisted initialization Then intensity and edge features are combined for a subsequent hybrid level-set segmentation. We evaluated our method on highly anisotropic clinical routine spine MRI datasets, containing 34 vertebrae, both healthy and pathological. We achieved a 3D Dice coefficient of 84.8\% and a mean surface-to-surface distance of 1.29 ± 0.42mm with regard to a manually created ground truth segmentation. The main advantages of our method are precise segmentation results on clinical routine images within reasonable processing time and with minimal user interaction.},
journal = {Procedia Computer Science},
author = {Hille, Georg and Glaßer, Sylvia and Tönnies, Klaus},
year = {2016},
keywords = {Clinical Spine MRI, Level-Sets, Precise, Segmentation, Vertebrae, Vertebral Bodies},
pages = {22 -- 27}
}