Development of a novel machine learning model to automate vertebral column segmentation utilizing biplanar full-body imaging.

Abstract

Degenerative scoliosis (DS) is a common spinal disorder among adults, characterized by lateral curvature of the spine. Recent advancements in biplanar full-body imaging, a low-dose and weight-bearing X-ray modality, facilitate safer and longitudinal imaging of DS patients. Quantifying spinal curvature serves as a valuable metric for assessing DS severity and informing surgical planning. However, manual annotation of vertebral structures in radiographic images is labor-intensive, necessitating specialized expertise and resulting in significant inter- and intraobserver variability. Advances in deep learning computer models, particularly with convolutional neural networks (CNNs) employing UNET architecture, offer robust solutions for image segmentation tasks. These deep learning approaches have the potential to standardize and expedite the analysis of spinal alignment alterations throughout disease progression. The purpose of this study is to develop an artificial intelligence algorithm capable of automating the segmentation of the vertebral column from biplanar full-body radiographic images regardless of spinal pathologies and previous hardware. This was a retrospective study designed to create and evaluate a proposed AI algorithm for spinal imaging. It was conducted in 2023 at a tertiary medical center and utilized weight-bearing, full-length biplanar full-body X-ray images in AP and Lateral orientations. The images were retrieved from the institutional picture archiving and communication system (PACS), anonymized, and exported as high resolution files. This study consisted of 250 images of patients who were either positive or negative for AIS. The primary outcome of this study was to identify the accuracy of the segmentation model using the Dice-Sørensen coefficient for anterior-posterior and lateral views. Biplanar full-body X-ray images were retrieved from the institutional picture archiving and communication system (PACS), anonymized, and exported as high-resolution files. Image dataset was crafted to include DS positive and negative samples. For each orientation, 200 images were used to train the model, and 50 radiographs were withheld for model performance evaluation. A two-stage deep learning model was developed to first identify the spine region from a full-body X-ray image, and then isolate the spine curvature from the output of the first stage of the model. The model was successful in segmenting the vertebral column, with Dice-Sørensen coefficient of 0.92 and 0.96 for anterior-posterior and lateral views respectively. The model was capable of accurately segmenting images involving complex spinal pathologies, such as lordosis and scoliosis, and noise from spinal instrumentation, such as rods and screws. Our findings indicate that a two-stage deep learning model with UNET architecture can accurately identify and segment spinal curvature in 2D biplanar full-body radiographs, offering a robust tool for DS assessment.