Autonomous extraction of preoperative radiographic predictors on X-ray for cervical spine deformity following laminoplasty: a prospectively validated AI tool.
Authors
Affiliations (7)
Affiliations (7)
- Department of Neurosurgery, Medical University of Gdansk, Gdansk, Poland.
- Neurosurgical Service, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, United States.
- Department of Computer Science, New York University, NYC, United States.
- Neurosurgical Service, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA, United States.
- Department of Neurosurgery, Medical University of Gdansk, Gdansk, Poland, Poland.
- Department of Neurosurgery, Pomeranian Medical University Hospital No. 1, Szczecin, Poland.
- Department of Neurosurgery, Pomeranian Medical University Hospital No. 1, Szczecin, Poland. [email protected].
Abstract
Approximately 21% of patients who undergo cervical laminoplasty for cervical spondylotic myelopathy (CSM) develop postoperative kyphotic deformity (KD). Radiologic parameters (RPs) on preoperative sagittal X-ray have consistently shown to be the strongest predictors for KD but their acquisition requires manual labor from specialists. Thus, the authors developed a novel artificial intelligence (AI) model to autonomously retrieve the predictors. A total of 259 patients' sagittal X-rays were retrospectively obtained from the internal center between 2016 and 2024 for training. Patients with spinal deformities, prior surgical interventions, or tumors/malignancies within the cervical region were excluded. Data augmentation techniques were used to amplify the dataset, and a custom attention U-net architecture was used. Prospective enrollment of patients diagnosed with CSM over nine months at the internal and external center was performed for validation. A total of 28 (77.8%) patients with CSM were prospectively obtained. The mean age was 66.0 ± 10.5 years, and 20 (71.4%) were females. The mean duration for human extraction of all RPs among the 2 neurosurgeons was 116.2 ± 17.5 seconds, whereas AI performed the task in 0.7 ± 0.0 seconds (p < 0.001). No significant difference was observed between the AI-obtained RPs and the physicians' report RP values, except for the center of gravity of the head to the C7 sagittal vertical axis, which reported a significant mean difference (p = 0.049); however, the extent of the difference was minimal (2.3 ± 4.0 mm). This study automated the extraction of RPs for postoperative KD following laminoplasty. The trained model is publicly available for software developers to implement on a standalone platform or as a plugin on a medical imaging viewer. The model thus incentivizes the development of a risk scoring system for KD that utilizes the AI-acquired RPs to improve the evidence-based practice when selecting the surgical approach.