Scholarlink

Objective <p>To evaluate the accuracy of a commercial deep learning algorithm in measuring spinopelvic parameters on full spine radiographs.</p> Materials and methods <p>This retrospective study analyzed total spine X-rays from a clinical cohort, assessing spinopelvic parameters (coronal Cobb angle, thoracic kyphosis, lumbar lordosis, pelvic incidence, sagittal vertical axis). Measurements were performed by clinical radiologists, a trained research reader serving as ground truth, and a commercial AI software. Inter-rater reliability was determined via intraclass correlation coefficients between an orthopedic surgeon and the trained reader (<i>n</i> = 50 images). Performance of both the AI and radiologists was statistically compared to the ground truth using mean absolute error, intraclass correlation coefficients, the Spearman correlation, and diagnostic accuracy metrics (sensitivity, specificity, predictive values).</p> Results <p>Four hundred ninety-five radiographs were analyzed. Inter-rater reliability between human raters, assessed via intraclass correlation coefficients, was excellent (0.94–1). Agreement between the algorithm and ground truth, also assessed via intraclass correlation coefficients, was good to excellent (0.79–0.99), except for lumbar lordosis (0.68, moderate). Mean absolute errors were lowest for coronal Cobb angles (4,6°; 95% confidence interval = 3.6–5.5°) and highest for lumbar lordosis (8.1°; 95% confidence interval = 6.9–9.3°). The Spearman rank correlations ranged from 0.74 to 0.99, and sensitivity was moderate to excellent (72.6–94.8). These results closely matched those obtained when comparing radiologists to ground truth. In a subgroup of patients with prior spinal fusion surgery, the correlation between the algorithm predictions and the ground truth was reduced, and measurement deviations were higher compared to non-instrumented patients.</p> Conclusion <p>The commercial AI software predicted most spinopelvic parameters with good reliability and accuracy, coinciding with radiologists in clinical practice, however, showed limitations in patients with spinal instrumentation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Diagnostic accuracy of artificial intelligence for spinopelvic parameters in standing total spine X-ray and limitations after fusion surgery

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月4日

查看全文
  • Tobias Winkler,
  • Thilo Khakzad,
  • Matthias Pumberger,
  • Friederike Schömig,
  • Matthias Becher,
  • Torsten Diekhoff

摘要

Objective

To evaluate the accuracy of a commercial deep learning algorithm in measuring spinopelvic parameters on full spine radiographs.

Materials and methods

This retrospective study analyzed total spine X-rays from a clinical cohort, assessing spinopelvic parameters (coronal Cobb angle, thoracic kyphosis, lumbar lordosis, pelvic incidence, sagittal vertical axis). Measurements were performed by clinical radiologists, a trained research reader serving as ground truth, and a commercial AI software. Inter-rater reliability was determined via intraclass correlation coefficients between an orthopedic surgeon and the trained reader (n = 50 images). Performance of both the AI and radiologists was statistically compared to the ground truth using mean absolute error, intraclass correlation coefficients, the Spearman correlation, and diagnostic accuracy metrics (sensitivity, specificity, predictive values).

Results

Four hundred ninety-five radiographs were analyzed. Inter-rater reliability between human raters, assessed via intraclass correlation coefficients, was excellent (0.94–1). Agreement between the algorithm and ground truth, also assessed via intraclass correlation coefficients, was good to excellent (0.79–0.99), except for lumbar lordosis (0.68, moderate). Mean absolute errors were lowest for coronal Cobb angles (4,6°; 95% confidence interval = 3.6–5.5°) and highest for lumbar lordosis (8.1°; 95% confidence interval = 6.9–9.3°). The Spearman rank correlations ranged from 0.74 to 0.99, and sensitivity was moderate to excellent (72.6–94.8). These results closely matched those obtained when comparing radiologists to ground truth. In a subgroup of patients with prior spinal fusion surgery, the correlation between the algorithm predictions and the ground truth was reduced, and measurement deviations were higher compared to non-instrumented patients.

Conclusion

The commercial AI software predicted most spinopelvic parameters with good reliability and accuracy, coinciding with radiologists in clinical practice, however, showed limitations in patients with spinal instrumentation.