A Multivariate Regression Model for Predicting the T2-T12 Kyphosis in Adolescent Idiopathic Scoliosis

Analyzing the sagittal profile of a scoliotic spine on the X-ray image can be at times quite difficult due to several anatomical structures overlapping in the thoracic region and irregular dosage of the X-ray beam which hinders the visibility of the vertebrae. There exists a known correlation between preoperative kyphosis and lordosis, as well as between the 2D curvature of the main coronal thoracic spine and the sagittal thoracic kyphosis. A predictor built on the most reliable information extracted from the X-ray images may therefore be able to estimate the thoracic kyphosis measurement between the T2 and T12 vertebrae. The main thoracic Cobb angle (CMT) taken on the coronal X-ray image as well as on the lumbar lordosis (LTL/L) computed on the sagittal image was used to built a reliable predictor of the constrained thoracic kyphosis angle between T2 and T12 (KT2-T12) by training a quadratic multi-variant regression model by cross-validation. A database containing 732 scoliotic spines demonstrating several types of scoliotic deformities was used to train the proposed system. The database was separated into two parts, with training and testing subset of 366 scoliotic spines each. The cross-validation accuracy (ratio of correctly predicted angle within 1°) computed from the training dataset on 366 scoliotic spines was of 0.98. The accuracy on the testing datasets was of 0.78 for hypo-kyphotic curves (40 spines, kyphosis range 0–10°), 0.89 for normal kyphotic curves (257 spines, kyphosis range 10–40°) and 0.76 for hyper-kyphotic curves (69 spines, kyphosis >40°). This paper presents promising results for a multivariate regression model which is able to predict the sagittal thoracic kyphosis between T2 and T12 built on the thoracic Cobb and lumbar lordosis measures. Although the estimate is not perfectly accurate, it falls within a reasonable range to assess adequately the sagittal curve of the spine.