Skip to content

Publications, 2025

Exploring Cage Subsidence Within the Lumbar Spine in LLIF Patients: A Machine Learning Segmentation Analysis

Jhair A. Colán, Akshay Sankar, Michael R. Kann, Samuel Adida, Regan M. Shanahan, Shovan Bhatia, Nicolás M. Kass, Nikhil Sharma, David J. McCarthy, Andrew D. Legarreta, Joseph S. Hudson, Griffin Hurt, Adam S. Kanter, Vincent J. Miele, Jacob T. Biehl, Edward G. Andrews

Neurosurgery, Vol. 71, Supplement 1, Apr 2025

Clinical Publisher doi:10.1227/neu.0000000000003360_1216 BibTeX

Abstract

INTRODUCTION: Interbody subsidence is a prevalent complication post-lateral lumbar interbody fusion(LLIF) among patients with low bone mineral density(BMD). While existing metrics evaluate BMD, none offer the granularity to assess individual vertebrae. Utilizing a machine learning(ML)-based automated computed tomography(CT) segmentation model, we can now achieve precise volumetric assessment of paraspinal musculature as well as Hounsfield(HU) distribution throughout the spine and individual vertebrae, serving as a proxy for BMD. METHODS: A retrospective analysis of 34 LLIF patients without posterior fusion (2014-2022) at a single institution was conducted. DICOM files were segmented using an open-source ML-program and analyzed using Python 3.9.6. Post-operative implant subsidence, graded according to Marchi criteria by two separate spine surgeons, was evaluated. Parameters included, but not limited to, average upper and lower vertebra(UV/LV HU) and endplates(EP-HU) HUs adjacent to cage, average trabecular bone HUs adjacent to cage(Tb-HU), average vertebral volume from L1-L5(cm3), and average lumbar paraspinal muscle volume(cm3). RESULTS: Twelve patients (35%) received PEEK implants, the remaining 22 received titanium implants. While all vertebral density parameters exhibited lower HU values in the subsidence cohort (n=6; Age 70.3±6.95; BMI 31.0±6.35; Surgical Levels=8) compared to the non-subsidence cohort (n=28; Age 70.3±6.95; BMI 31.0±6.35;Surgical Levels=52), only endplate HUs was significantly correlated to subsidence (EP 444.2±110.44vs510.6±116.03, p=0.036). Volumetric parameters also demonstrated lower cm3 values in the subsidence group versus the non-subsidence group (Lumbar Vertebral Volume 38.4±10.30vs40.7±11.22,p=0.644; Paraspinal Volume 335.4±79.87vs388.1±110.62,p=0.364). CONCLUSIONS: Automated ML segmentation identified lower endplate HU as significantly correlated to cage subsidence following LLIF. Future pre-operative HU analysis may help avoid this common LLIF complication.

BibTeX

@article{colan2025cage,
  title={Exploring Cage Subsidence Within the Lumbar Spine in LLIF Patients: A Machine Learning Segmentation Analysis},
  author={Colán, Jhair A. and Sankar, Akshay and Kann, Michael R. and Adida, Samuel and Shanahan, Regan M. and Bhatia, Shovan and Kass, Nicolás M. and Sharma, Nikhil and McCarthy, David J. and Legarreta, Andrew D. and Hudson, Joseph S. and Hurt, Griffin and Kanter, Adam S. and Miele, Vincent J. and Biehl, Jacob T. and Andrews, Edward G.},
  journal={Neurosurgery},
  volume={71},
  number={Supplement 1},
  pages={192},
  year={2025},
  month=apr,
  publisher={Ovid Technologies (Wolters Kluwer Health)},
  doi={10.1227/neu.0000000000003360_1216},
}