Spine Research Fellow University at Buffalo Neurosurgery (UBNS) University at Buffalo Neurosurgery (UBNS) Buffalo, NY, US
Disclosure(s):
Jacob D. Greisman, MD: No financial relationships to disclose
Introduction: Accurate measurement of spinal parameters is essential for diagnosing and treating spinal disorders. Traditional manual measurement methods are time-consuming and prone to variability. This study aims to assess the accuracy and reliability of machine learning (ML) methods compared to traditional measurements performed by a neurosurgery resident (operator 1).
Methods: We conducted a comparative analysis of spinal parameter measurements obtained through ML models and those manually taken. Parameters examined included cervical and thoracic kyphosis, lumbar lordosis, pelvic incidence (PI), sacral slope (SS), pelvic tilt (PT), sagittal vertical axis (SVA), C7 coronal alignment (C7 CSL), thoracic Cobb angle, and lumbar Cobb angle. Data analysis utilized Pearson coefficients, p-values, and interrater reliability.
Results: The results demonstrated a high correlation and significant agreement between measurements obtained through ML and Operator 1, especially in the lumbar spine. Sacral slope (SS) and pelvic tilt (PT) demonstrated Pearson coefficients of 0.84 and 0.90, respectively (both with p-values < 0.001), and interrater reliabilities of 0.84 and 0.90, respectively. The sagittal vertical axis (SVA) exhibited a Pearson coefficient of 0.96 (p-value < 0.001) with an interrater reliability of 0.96. PI-LL and PI had Pearson coefficients of 0.59 and 0.79, respectively (p-values < 0.001), and interrater reliabilities of 0.52 and 0.79, respectively.
Kyphosis C2/C7 exhibited a Pearson coefficient of 0.68 (p-value < 0.001) with an interrater reliability of 0.66. Similarly, lordosis C2/C7 showed a Pearson coefficient of 0.58 (p-value < 0.001) with an interrater reliability of 0.54. Kyphosis T4-T12 demonstrated a Pearson coefficient of 0.79 (p-value < 0.001) with an interrater reliability of 0.76. However, C7 CSL had a lower agreement, with a Pearson coefficient of 0.32 (p-value < 0.001) and an interrater reliability of 0.20. Thoracic and lumbar Cobb angles showed Pearson coefficients of 0.66 (p-value < 0.001) and interrater reliabilities of 0.55 and 0.70, respectively.
Conclusion : Machine learning facilitates the accuracy and reliability of measuring lumbar spinal parameters. However, future research should explore challenges in coronal parameters.