Document Type

Article

Publication Title

Artificial Intelligence Surgery

Abstract

Aim: Quantitative measurement of spinopelvic parameters from radiographs is important for assessing spinal disorders but is limited by the subjectivity and inefficiency of manual techniques. Deep learning may enable automated measurement with accuracy rivaling human readers.

Methods: PubMed, Embase, Scopus, and Cochrane databases were searched for relevant studies. Eligible studies were published in English, used deep learning for automated spinopelvic measurement from radiographs, and reported performance against human raters. Mean absolute errors and correlation coefficients were pooled in a meta-analysis.

Results: Fifteen studies analyzing over 10,000 radiographs met the inclusion criteria, employing convolutional neural networks (CNNs) and other deep learning architectures. Pooled mean absolute errors were 4.3° [95% confidence interval (CI) 3.2-5.4] for Cobb angle, 3.9° (95%CI 2.7-5.1) for thoracic kyphosis, 3.6° (95%CI 2.8-4.4) for lumbar lordosis, 1.9° (95%CI 1.3-2.5) for pelvic tilt (PT), 4.1° (95%CI 2.7-5.5) for pelvic incidence (PI), and 1.3 cm (95%CI 0.9-1.7) for sagittal vertical axis (SVA). Intraclass correlation coefficients exceeded 0.81, indicating strong agreement between automated and manual measurements.

Conclusion: Deep learning demonstrates promising accuracy for automated spinopelvic measurement, potentially rivaling experienced human readers. However, further optimization and rigorous multicenter validation are required before clinical implementation. These technologies may eventually improve the efficiency and reliability of quantitative spine image analysis.

DOI

10.20517/ais.2024.36

Publication Date

1-3-2025

Keywords

Deep learning, spine parameters, pelvic parameters

ISSN

2771-0408

Download

Share

COinS