Deep learning-based estimation of sagittal spinal alignment from coronal radiographs

Area of Science:

• Spine biomechanics
• Medical imaging analysis
• Deep learning applications in radiology

Background:

• Patient-specific sagittal alignment parameters are crucial for spinal biomechanical modeling.
• Traditional biplanar radiographs involve radiation exposure, prompting research into alternative methods.
• Previous deep learning attempts for cross-plane prediction, like GANs, have not met clinical accuracy standards.

Purpose of the Study:

• To investigate multiple-output regression (MOR) for directly predicting sagittal spinal parameters from coronal radiographs.
• To evaluate the performance of MOR against generative adversarial network (GAN)-based approaches for this task.

Main Methods:

• A convolutional neural network-based MOR model was trained and tested on a dataset of 449 adult subjects' EOS biplanar radiographs.
• Six key sagittal parameters (TK, LL, SS, PI, PT, SVA) were predicted from coronal images.
• Performance was assessed using absolute error, regression analysis, Bland-Altman plots, and Lin's concordance correlation coefficient (CCC), with external validation on an adolescent idiopathic scoliosis (AIS) dataset.

Main Results:

• In adults, MOR showed median absolute errors of 4°-8° for angular parameters and 1.6 cm for SVA, with significant biases.
• Concordance correlation coefficients (CCC) ranged from 0.20 to 0.69, indicating weak to moderate agreement for most parameters.
• In AIS patients, MOR and GANs had comparable median errors (~7°), though GANs performed better for SVA (1 cm vs. 2 cm).

Conclusions:

• Direct MOR prediction of sagittal parameters from coronal radiographs yields limited accuracy and systematic biases.
• The comparable performance of MOR and GANs suggests insufficient sagittal information in coronal views is the main limitation.
• Current deep learning techniques are inadequate to replace biplanar imaging for reliable spinal sagittal assessment.