Logo image
Back
Finite element and neural network-based modeling for simulation of scoliosis progression and correction
Dissertation

Finite element and neural network-based modeling for simulation of scoliosis progression and correction

Christian Reynold D'Andrea
Doctor of Philosophy (Ph.D.), Drexel University
Aug 2026
DOI:
https://doi.org/10.17918/00011580
pdf
DAndrea_Christian_2026120.68 MB
PDF Embargoed Access, Embargo ends: 30 Sep 2028

Abstract

Finite elements Growth modulation Neural networks Scoliosis Spine biomechanics Surgical planning Computational Physics
Spine fusion is the standard surgical treatment for correction of severe adolescent scoliosis, but it suppresses remaining growth. Growth-modulating approaches such as vertebral body tethering (VBT) have been developed to mitigate fusion downsides, acting through both on-table correction and sustained asymmetric mechanical alteration of vertebral growth to correct curvature over years. These procedures present a planning challenge by requiring a surgeon to anticipate the correction occurring post-operatively due to years of stress-altered growth. Incomplete curve correction and implant breakage have been observed, and a simulation tool may contribute to improved outcomes by assisting surgeons in planning such procedures, predicting long-term growth-based correction while accounting for patient-specific factors including flexibility and skeletal maturity. This work laid a foundation for such a simulation tool: FE models personalized to the detailed spine geometry of 20 VBT patients and calibrated to surgical and subsequent stress-modulated growth correction were created, achieving a median final thoracic Cobb-angle error of 3.4°. Beyond VBT, new growth-modulating implants are being developed, and are tested using expensive and time-consuming porcine in vivo testing. To support preclinical screening, porcine-specific FE models were calibrated to serial imaging after implantation of a novel tether-based device. Across three pigs, the models reproduced frontal-plane deformity over 16 weeks with Cobb-angle mean absolute errors of 4.6-8.5°. Conventional FE simulations may require hours per solution, precluding adaptive real-time intraoperative planning. Neural-network-based methods have been demonstrated to learn an FE model's solutions, but still often require thousands of conventional FE evaluations for training. Training neural nets directly using material constitutive laws and without use of conventional FE solutions may enable faster training, but this approach is not well-explored in biomechanics. The approach was applied to intervertebral disc deformation and lumbar spine bending. Within tested geometries and loads, the IVD model reproduced matched FE behavior, while the lumbar model achieved range-of-motion errors below 4°. For a vertebral-column-only benchmark, post-compilation forward evaluation and separately measured field recovery totaled approximately 0.5 seconds, compared with approximately 30 minutes for FE. Taken together, these studies establish foundations for preclinical implant evaluation, patient-specific preoperative planning, and rapid spine analyses with future intraoperative potential for pediatric scoliosis correction through growth modulation.

Metrics

2 Record Views

Details

Logo image