Direct Prediction of Sagittal Spinal Parameters from Coronal Radiographs Remains Inaccurate

A deep learning model directly predicting sagittal spinal alignment from coronal radiographs achieved only weak-to-moderate agreement (CCC 0.20-0.69) in adults, with median absolute errors of 4°-8°. Performance was similar to a GAN-based approach, indicating a fundamental limita…

• In the adult test set, median absolute errors ranged from 4° to 8° for angular parameters, but maximum errors reached up to 40° for angles and 8.8 cm for sagittal vertical axis (SVA), indicating wide variability.
• External validation in 69 adolescents with idiopathic scoliosis showed comparable performance between the regression model and a GAN-based method (median errors ~7°), though the GAN was more accurate for SVA (1 cm vs. 2 cm).
• The study concludes that coronal radiographs inherently lack sufficient information for reliable sagittal assessment, and current deep learning methods are inadequate to replace biplanar imaging.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.