Introduction. The analysis of the spinal alignment in the standing posture has gained wide attention in the last years. Among the large available body of literature, several studies investigated the spinal alignment of cohorts of patients with adult spine deformities (ASD) as well as of young subjects suffering from adolescent idiopathic scoliosis (AIS). Biplanar radiography is the state-of-the-art imaging technique for such investigations, but its use in large studies is restricted by the substantial manual labour required for the three-dimensional reconstruction of the spinal anatomy necessary to calculate the radiographic parameters. Therefore, most studies tend to be limited to relatively small cohorts. The aim of this study is to analyze a large dataset of biplanar radiographs of 9823 non-operated subjects by employing an automated tool based on artificial intelligence, rather the labour-intensive manual reconstruction.

Methods. A deep learning model based on a convolutional neural network using a Differentiable Spatial to Numerical Transform (DSNT) top layer has been developed and used to determine the coordinates of 10 vertebral landmarks in the T1-sacrum region, as well as the hip centers. The model was used to process biplanar radiographs in standing of the full trunk of 9823 consecutive patients acquired between 2015 and 2018 and stored in the local imaging database of the institute. The coordinates of the landmarks were then used to calculate the value of several radiographic parameters (Cobb angle of scoliosis, spinopelvic parameters, lumbar lordosis, thoracic kyphosis, sagittal vertical axis, apex rotation in the transverse plane) and to perform a descriptive statistics analysis after stratifying the subjects based on the type of deformity (no deformity, AIS, ASD) and on age groups.

Results. Among the 9823 patients, 3018 suffered from AIS, 2865 from ASD, and 3940 had no spinal deformities. The descriptive statistical analysis highlighted several findings not observed in previous smaller studies; for example, it could be shown that the pelvic incidence-lumbar lordosis mismatch clearly depends on the pelvic incidence itself regardless of the type and severity of spinal deformity, and that the lumbar lordosis decreases with increasing Cobb angles in both AIS and ASD patients, but more markedly in the latter case (Figure).

Discussion. The deep learning model allowed analyzing a very large database of biplanar spine radiographs, permitting for the first time a large-scale three-dimensional investigation of spinal deformities. Among the limitations of the study, it should be noted that information about the clinical indication for imaging was not available, and that the images were acquired in a raised-arm posture which may influence the radiographic parameters. In summary, the use of deep learning for large-scale studies has the potential to reveal clinically relevant information that does not emerge when smaller cohorts are investigated.

Figure caption. Biplanar radiographs of two exemplary patients with the corresponding vertebral landmarks detected by the automated deep learning tool (top); scatter and regression plots showing the correlations between pelvic incidence and the pelvic incidence–lumbar lordosis mismatch (bottom left) and those between the Cobb angle of scoliosis and the lumbar lordosis (bottom right).