A machine learning based surgical system for optimized spine surgery in patients with osteoporosis

Abstract

The local bone mineral density and bone microarchitectures in different vertebral subregions change unevenly for patients with osteoporosis. Computed tomography (CT) image-based artificial intelligence (AI) techniques are widely studied in the imaging biomarker field, but research on applying such techniques in osteoporotic fractures and surgical planning is limited. This thesis aimed to develop an AI- empowered system to accurately assess local bone quality and to generate an optimal surgical plan for patients with osteoporotic spinal fractures.

In the first study, morphological operation residual blocks were used to improve morphological feature representation in convolutional neural networks for semantic segmentation of medical images. In end-to-end deep learning models for semantic segmentation, the 3D morphological operation residual block extracts morphological characteristics. Experimental results demonstrated that the proposed network performed better than the conventional approach in segmentation tasks.

In the second study, a routine CT-based measurement was explored for focal osteoporosis defect quantification, and its association with the vertebral compression fracture prevalence was investigated. A total of 205 cases were retrospectively reviewed and enrolled into either the vertebral compression fracture or the control group. The measurement of focal Bone Mineral Content (BMC) loss showed high reproducibility (RMSSD = 0.011mm, LSC = 0.030mm, ICC = 0.97), and high correlation with focal bone volume fraction (r = 0.79, P<0.001) and trabecular bone separation (r = 0.76, P<0.001), but poor correlation with trabecular bone mineral density (BMD) (r = 0.37, P<0.001). The focal BMC loss was significantly higher in the fracture group than in the control group (1.03±0.13 vs. 0.93±0.11 mm; P<0.001) and was associated with prevalent vertebral compression fracture (1.87, 95% CI = 1.31~2.65, P<0.001) independent of trabecular BMD.

In the third study, subregional vertebral BMD was investigated to measure local bone quality. Quantitative CT images of 115 people (62 women, 53 men; mean age = 66.4±13.4 years) were retrospectively collected, from which we manually segmented 403 lumbar vertebral bodies. This automatic approach achieved high segmentation performance for vertebral body segmentation (accuracy 0.98±0.02, dice coefficient 0.92±0.06, IoU 0.87 0.09), cortical bone segmentation (accuracy 0.95±0.02, dice coefficient 0.92±0.03, IoU 0.85 0.05), and endplate segmentation (accuracy 0.89±0.05, dice coefficient 0.75±0.09, IoU 0.61 0.12). Through the segmentation algorithm of vertebral body subregions, a deep learning model was generated that ensures high local BMD measurement precision and reproducibility.

Finally, a novel AI-empowered surgical planning system was explored using a local bone quality assessment tool for trajectory planning of pedicle screw and interbody cage. The preoperative CT scans of 21 elderly osteoporotic patients (69.6±7.8 years old, 55.9±17.1mg/cc BMD) were retrospectively analysed from multi-centre hospitals. On both sides of pedicles in L3-L5, the optimized trajectories of pedicle screw showed significantly higher BMD and pull-out strength than those of standard trajectories referenced by the AO foundation (p<0.05) in 126 pedicles with at least 2.0-fold increase. In addition, all different positions and angles for cage placement were simulated. The BMD in the best position of the endplate under the cage showed at least a 31.39% increase compared with that in the worst position of cage.