An existing tool able to automatically perform a 3D reconstruction of the anatomy and shape of the thoracolumbar spine based on biplanar radiographs acquired with the EOS Imaging System has been used as the basis for the current study (Galbusera et al., 2019). The deep learning model has been extended by increasing the number of recognized vertebral landmarks from 2 to 10 (Figure 1) and implementing the self-supervision paradigm to improve its performance. In particular, using 10 landmarks allowed for a greater precision in the calculation of the endplate orientation with respect to the previous version, in which we simply considered the line orthogonal to the spline interpolating the two endplate centers. The self-supervised approach allowed exploiting also a set of unlabelled images in the training process in order to increase the model understanding on images. In detail, following the approach presented in Honari et al. (2018) the model produced landmarks localizations equivariant with respect to a set of transformations (rotations and translations) applied to the image. So, for the unlabelled images, if a transformation is applied to an image, the transformed landmarks should be very close to the points obtained by applying the same transformation to landmark coordinates in the original image. In brief, g(T(I)) ≈ T(g(I)) where g is the network, T the transformation and I the image. As regards the model architecture, while the Differentiable Spatial to Numerical Transform (DSNT) top layer for the regression of the landmark coordinates was kept unaltered from the previous version, the backbone was changed from a simple 7-layer convolutional neural network to a ResNet-34. The size of the training set was extended from 443 to 810.

The accuracy of the tool was quantitatively assessed by comparing the automated prediction of various relevant radiological angles with measurements performed by four human observers with an in-house computer-aided tool. The comparison was conducted on three sets of biplanar images of 30 patients which were not included in the training data. The first set referred to patients with no spinal deformity (age 25–70), the second set to patients with adolescent idiopathic scoliosis (age 11–18, coronal Cobb angle 20–65°), while the third set included patients suffering from adult spinal deformities (age 66–85, coronal Cobb angle 3–80°, SVA 1–12 cm). To validate the tool, we counted the number of cases for which the parameter values estimated by the tool fell within the range of the measurements of the four observers, as well as those with an error greater than 5°.

A large set of biplanar radiographs of the full trunk or whole body of 16,228 patients consecutively acquired between 2015 and 2019 at IRCCS Istituto Ortopedico Galeazzi was retrospectively analyzed (Figure 1). All images were acquired in standing in a raised-arm posture to allow for an optimal visualization of the spine in both planes. Non-relevant images were excluded based on the following criteria: 1) age of the subject below 10 years; 2) presence of spinal implants in the images or evidence of non-instrumented spine surgery; 3) neuromuscular and congenital spinal deformities. Patients were stratified based on sex, age group (10–18 years old; 19–44; 45–64; 65–79; 80 or more) as well as on the presence of a spinal deformity, which was defined as a maximal Cobb angle in the coronal plane in the thoracolumbar spine greater than 10° (Lonstein, 1994) and/or a sagittal vertical axis (SVA) greater than 5 cm (Glassman et al., 2005).

For all patients, after using the deep learning model for the prediction of the three-dimensional position of the vertebral landmarks, the values of the following radiological parameters were automatically calculated: spinopelvic parameters (pelvic incidence (PI), pelvic tilt (PT), sacral slope (SS)); lumbar lordosis between L1 and L5 (LL) as well as based on Roussouly’s definition (Roussouly et al., 2003); mismatch between pelvic incidence and lumbar lordosis; thoracic kyphosis between T1 and T12 and between T4 and T12 (TK); maximal Cobb angle in the coronal plane between T1 and L5; SVA; T1-pelvic angle (TPA); in case of scoliosis, maximal vertebral rotation in the transverse plane.

Scatter plots representing potential correlations of clinical interest between the radiographic parameters were built; linear regression analysis was performed for the same correlations. Analysis of covariance (ANCOVA) was employed to test differences between sexes and age groups in terms of LL, TK, PI, SVA and pelvic incidence-lumbar lordosis mismatch (PI-LL), considering various covariates (age, maximal coronal Cobb angle, PI, LL). Besides, the matrix describing the pairwise Spearman correlation coefficients between various relevant parameters (age, maximal coronal Cobb angle, PI, SS, LL, TK, SVA, TPA, maximal axial rotation) was computed. Finally, the importance of some relevant demographic and radiological parameters (age, sex, PI, SS, LL, TK, SVA, maximal coronal Cobb angle, maximal axial rotation) in determining parameters describing the spinal alignment and possible compensatory mechanisms (LL, TK, SVA, PI-LL, TPA, maximal coronal Cobb angle, maximal axial rotation) were determined by means of a gradient boosted decision tree implemented with the XGBoost Python library (https://xgboost.readthedocs.io/).

On average, in 43% of the cases (range 31–72%) the radiological parameters predicted by the automated tool were within the range of the measurements by the four human observers (Figure 2). The cases with error larger than 5° ranged between 0% (pelvic tilt in patients with adult spine deformity) and 20% (maximal coronal Cobb angle in patients with adolescent idiopathic scoliosis).

After the application of the exclusion criteria, biplanar images of 9,832 patients were included in the study. Among these, 5,122 patients had an age between 10 and 18 years, 1,246 between 19 and 44 years, 1,645 between 45 and 64, 1,579 between 65 and 79, and 240 higher than 80 years 5,006 patients were found to have scoliosis, i.e., a maximal coronal Cobb angle higher than 10°, 2,710 of them belonging to the 10–18 years age group. Among the 9,832 patients, 4,428 were found to have no clinically relevant spine deformities, i.e., Cobb angle lower than 10° and SVA lower than 5 cm. Table 1 shows range, mean and median values of several radiological parameters within the population.

The radiological parameters describing the sagittal alignment of the spine were found to be highly interrelated with each other, as well as dependent on age, while sex had relatively minor importance (Figure 3). Nevertheless, all performed ANCOVA analyses to test differences between age groups as well as sex in terms of sagittal parameters showed statistical significance regardless of the magnitude of such differences, due to high numerosity of the sample. The full set of charts describing the associations between the radiological parameters is reported in the Supplementary Material.

The L1-L5 lumbar lordosis was associated with both the thoracic kyphosis and SVA (Figure 4). While the increase in thoracic kyphosis with respect to the change in lumbar lordosis, i.e., the slope of the regression curve, was weakly dependent on the age, older subjects had markedly higher kyphosis with respect to younger ones being the lumbar lordosis equal. In contrast, the slope of the regression curve describing SVA with respect to the lumbar lordosis was dependent on the age, i.e., older subjects had higher changes in SVA being the change in lumbar lordosis equal, demonstrating the importance of age in determining how the global sagittal balance responds to a decrease in lumbar lordosis.

A strong association between PI and lumbar lordosis was found, while the thoracic kyphosis showed a less clear correlation with PI (Figure 5). Regarding the lumbar lordosis, while all age groups showed a positive correlation, the slope of the regression lines was higher for younger subjects, with minimal differences between the 10–18 and 19–44 age groups as well as for subjects older than 65 years. The thoracic kyphosis tended to decrease with increasing PI, with a greater effect for the younger subjects.

The pelvic incidence-lumbar lordosis mismatch was found to be dependent on PI, as well as on age (Figure 6). In general, PI-LL values increased with increasing PI, with the slope of the regression line showing only minor differences among the age groups. The same trend was observed even after excluding subjects with spinal deformities.

The severity of scoliosis, as described by the maximal Cobb angle in the coronal plane, had a distinct association with the sagittal alignment in adolescent subjects with respect to the older patients (Figure 7). Similar to the sagittal parameters, all ANCOVA analyses (i.e. effect of sex and age group considering the coronal Cobb angle as covariate) showed statistical significance, even if the magnitude of such effects was small. While the Cobb angle was weakly associated with a change in lumbar lordosis in the group of young subjects, the lordosis showed a marked decrease for more severe curves in adults. On the contrary, the thoracic kyphosis showed a clear tendency toward a decrease with the severity of the scoliotic curve in both adolescents and adults, with approximately similar behaviours.

The matrix showing the Spearman correlation coefficients between various parameters confirmed the strong interdependence between them, as well as the findings widely documented in the literature and highlighted in the previous paragraphs (Figure 8). Higher age is associated with lower LL, higher TK, SVA and TPA, as well pI. Higher PI is correlated with higher SS, LL, SVA and TPA, whereas it had a minor correlation with TK. SS and LL show a high correlation, as well as the maximal Cobb angle and the maximal rotation in the axial plane.

The XGBoost model showed that the importance of the demographic and radiological parameters in determining the spinal alignment did not exhibit major differences between adolescent and adult subjects (Figure 9). The spinopelvic parameters, especially the sacral slope which includes both anatomical and postural information, together with the lumbar lordosis had as expected high importance in determining other sagittal parameters. In adults, SVA (that describes the global sagittal balance) was mostly determined by the lumbar lordosis, while in adolescents all sagittal parameters contributed to it. The strongest predictor for the coronal Cobb angle was the maximal rotation in the transverse plane, while no radiological parameter describing the sagittal alignment (with the partial exception of the thoracic kyphosis for the adolescent subjects) seemed to play a major role in its determination.