Deep Learning Automates Measurement of Spinopelvic Parameters on Lateral Lumbar Radiographs

John T. Schwartz, Brian H. Cho, Peter Tang, Javin Schefflein, Varun Arvind, Jun S. Kim, Amish H. Doshi, Samuel K. Cho

Research output: Contribution to journalArticlepeer-review

27 Scopus citations

Abstract

Study Design.Cross-sectional database study.Objective.The objective of this study was to develop an algorithm for the automated measurement of spinopelvic parameters on lateral lumbar radiographs with comparable accuracy to surgeons.Summary of Background Data.Sagittal alignment measurements are important for the evaluation of spinal disorders. Manual measurement methods are time-consuming and subject to rater-dependent error. Thus, a need exists to develop automated methods for obtaining sagittal measurements. Previous studies of automated measurement have been limited in accuracy, inapplicable to common plain films, or unable to measure pelvic parameters.Methods.Images from 816 patients receiving lateral lumbar radiographs were collected sequentially and used to develop a convolutional neural network (CNN) segmentation algorithm. A total of 653 (80%) of these radiographs were used to train and validate the CNN. This CNN was combined with a computer vision algorithm to create a pipeline for the fully automated measurement of spinopelvic parameters from lateral lumbar radiographs. The remaining 163 (20%) of radiographs were used to test this pipeline. Forty radiographs were selected from the test set and manually measured by three surgeons for comparison.Results.The CNN achieved an area under the receiver-operating curve of 0.956. Algorithm measurements of L1-S1 cobb angle, pelvic incidence, pelvic tilt, and sacral slope were not significantly different from surgeon measurement. In comparison to criterion standard measurement, the algorithm performed with a similar mean absolute difference to spine surgeons for L1-S1 Cobb angle (4.30°±4.14° vs. 4.99°±5.34°), pelvic tilt (2.14°±6.29° vs. 1.58°±5.97°), pelvic incidence (4.56°±5.40° vs. 3.74°±2.89°), and sacral slope (4.76°±6.93° vs. 4.75°±5.71°).Conclusion.This algorithm measures spinopelvic parameters on lateral lumbar radiographs with comparable accuracy to surgeons. The algorithm could be used to streamline clinical workflow or perform large scale studies of spinopelvic parameters.Level of Evidence: 3.

Original languageEnglish
Pages (from-to)E671-E678
JournalSpine
Volume46
Issue number12
DOIs
StatePublished - 15 Jun 2021

Keywords

  • Automation
  • artificial intelligence
  • interrater reliability
  • machine learning
  • pelvic incidence
  • pelvic tilt
  • radiography
  • sacral slope
  • sagittal alignment
  • spinopelvic measurement

Fingerprint

Dive into the research topics of 'Deep Learning Automates Measurement of Spinopelvic Parameters on Lateral Lumbar Radiographs'. Together they form a unique fingerprint.

Cite this