3D Markerless asymmetry analysis in the management of adolescent idiopathic scoliosis

Adolescent idiopathic scoliosis (AIS) is the most common form of three-dimensional (3D) spinal deformity. It affects 2–4% of the population, predominantly females [1]. The AIS spine deformity progresses rapidly during the adolescent growth period, resulting in a need for frequent follow-ups [2]. The gold standard for assessing the spine curve is measuring the Cobb angle on the full torso radiograph, defined as the angle between the two most tilted vertebrae in each curve [3].

The conventional monitoring of the scoliosis using the Cobb Angle has limitations recognized in the literature. Firstly, the measurement is limited to 2D posterior-anterior radiographs, and thus the method fails to address the 3D characteristics of AIS [4]. In addition, the use of radiographs in scoliosis clinics has several pitfalls of growing concern, such as excessive X-ray radiation exposure, with their associated risk of developing cancers [5‐9], and the contra-indication of radiograph acquisitions for pregnant women.

Surface topography (ST) was introduced as a new approach to improve the monitoring of the scoliosis [5, 10‐12]. ST is a method for which non-invasive visible light is used for scanning the torso surface in order to assess cosmetic deformities often based on some landmarks placed on the patient’s torso by trained clinicians and using related indices based on the coordinates of the landmarks with respect to each other and the geometric properties of transverse cross section of the torso, such as the cosmetic score and Quantec spinal angle [13‐17]. Measurements based on ST could possibly be used effectively in combination with the radiographs to decrease the radiation dose and risk of cancer resulting from the multiple X-ray acquisitions. At this point, the ST approach is by no means designed to fully replace the gold-standard radiograph measurements, because it is subject to validation with the radiograph measurements. Nevertheless, the development of accurate ST methods has significantly contributed to the management of scoliosis [12].

On the other hand, marker placements can be associated with human errors in collecting the raw data [14] and such methods fail to take into account the whole torso geometry in the analysis. In contrast, our team has developed a novel markerless ST asymmetry analysis approach, which is independent of human interactions, and considers the full 3D torso surface for the analysis [12]. This technique provides a deviation contour map that visualizes the areas affected by AIS corresponding to the location of each curve, called deviation patch [12]. These patches are isolated to calculate ST parameters for such deviations. The method demonstrated the potential for reducing 44% of radiograph exposure in the monitoring of scoliosis [15].

In the study presented by Komeili et al. [12], areas with deviation less than 3 mm were considered normal while greater deviations were considered as a deformation and separated the deviation patches for further analysis. Further application of this proposed ST asymmetry analysis in over 250 AIS patients illustrated that in 30 cases the method failed to either locate the curve properly or to correlate to the corresponding scoliotic curve, and led to misclassification of the AIS severity or progression. For example, in the ST analysis of some patients with a double scoliotic curve, a single isolated deviation patch encompassed the entire back torso (see Fig. 1), and therefore did not reflect the double scoliotic curve in the corresponding radiograph. In some other cases, the deviation patch extended to the anterior part of the torso due to the asymmetry introduced by the breasts or axial rotation of the torso. Folded skin near the armpits and waist also introduced artifact in the ST analysis. The reasons for such lack of correspondence between the surface and the radiographic results were traced to the patch isolating stage where 3 mm was used as a “cut point”. So far, these cases have been manually handled case by case by a scoliosis professional, which can introduce human errors in the measurements and decrease the correlation between the ST parameters and radiograph measurements.

This study aimed to eliminate patch overlaps by enhancing the patch isolation procedure, and to increase the accuracy of the analysis in identifying patients where we could prevent unnecessary X-ray exposure (mild patients and those who did not experience any progression from the last visit). Our hypothesis is that the modifications suggested in this study would isolate the deviation patches without compromising the accuracy of the method in monitoring the scoliotic curve severity or progression.

The markerless asymmetry analysis developed previously [12] showed promising results in identifying curves with progression in the follow-up visits with the potential ability of reducing 44% of radiation exposure in the monitoring of AIS. The research presented in this study aimed at modifying the markerless asymmetry analysis by simplifying the application of the previously developed method while also improving the sensitivity and accuracy of the prediction in identifying the moderate/severe curves and the progression of the scoliotic curve. While the need for manual handling of deviation patches was resolved and the sensitivity in identifying the progressed curves improved by 10%, the overall accuracy of the method decreased by 5%. As a result, the applied modifications on the original method resulted in more conservative monitoring and curve severity assessment of scoliosis where fewer cases of moderate/severe curves and progression cases were missed (see Fig. 5).

In this study, the optimum threshold of deviation for isolating the deviation patches was found to be 9.33 mm. The isolated deviation patches were generally smaller in size than those obtained in Komeili et al. [13, 19] (see Fig. 1). The modified code successfully prevented the extensions of isolated deviation patch to the anterior section of the torso and around the armpits, and clearly separated the boundaries between the deviation patches in ST scans of patients with a double or triple scoliotic curves shown in Fig. 1. Moreover, we expect that the reliability of the method in repeating the analysis would increase by automating the patch isolation step and avoiding human intervention in the ST analysis process.

Accuracies of 65 and 60% were obtained in categorizing the severity of T-TL and L curves in the Validation cohort, respectively as show in Fig. 3. This level of accuracy was approximately the same as the accuracy obtained for the Training group, indicating repeatability of the classification model in the prediction of T-TL curves severity.

The distribution of MaxDev and RMS parameters of subjects in the Training group in Fig. 4 illustrated that the correlation of ST parameters with the curve severity in the T-TL region was better than in the L region, which is because of the fact that the soft tissue in the L area dampens the deformation of the vertebrae and might soften the effect of the curve in the surface of the torso, while for T-TL curves, the curvature of the vertebrae is transferred to the back surface of the torso through the rib cage and results in prominent deformation [16]. All ST scans with MaxDev less than 9.33 mm in Fig. 4, which were denoted by ◊ or ♦, had Cobb angle less than 15 degrees and were correctly classified using the original method developed by Komeili et al. [12].

The monitoring of T-TL and L curves progression were analyzed separately (results are not shown) and no difference in the sensitivity, specificity, and accuracy was obtained if T-TL and L curves were analyzed in one group. Therefore, for the monitoring of scoliotic curve, T-TL and L curves were mixed and analyzed using only one classification tree which is likely more practical clinically (Fig. 6). The variation of ST parameters in this study was correlated with the progression of the Cobb angle measured in radiographs with a sensitivity of 67% for the Validation group in Fig. 6. However, some negative changes in ΔRMS and ΔMaxDev were obtained for patients with a positive ΔCA in Fig. 7. The proposed threshold of 9.33 mm for isolating the deviation patches in this study resulted in smaller accuracy and specificity in monitoring curve progression specifically for patients with mild and moderate curves in the baseline. Excluding areas with the deviation less than 9.33 mm from the deviation patches provided less information about the torso shape for monitoring the deformities over time. Therefore, any change in the areas with the deviation less than 9.33 mm over time was not included in the analysis, which led to the misclassification of multiple curves especially in the patients with moderate scoliotic curves, where relatively small portion of torso had deviations greater than 9.33 mm (see Fig. 7). Our results showed that curve progression monitoring appears to be 100% (4/4) accurate for patients with severe scoliotic curve in baseline who progressed in the follow-up scans (tabulated results are not shown). Therefore, monitoring mild and moderate scoliotic curves that present smaller asymmetry on the torso surface with respect to the severe curves, using the modified asymmetry analysis method in this study involves a higher risk of false positive. The case-by-case investigation of patients who were incorrectly identified in the Progression group (false positive) showed that 55% of the population had truly a larger Cobb angle in the follow-up radiographs compared to the baseline. However, the positive increase of Cobb angle was not significant enough to be considered as a progression, i.e., ΔCA in the range of 0–5 degrees. Although a follow-up radiograph is not necessary for this group of patients, considering the error in Cobb angle measurements and the small degree of progression in many of the participants might sufficiently justify taking a radiograph in the follow-up.

The overall sensitivities in identifying curve severity in this study were similar to the corresponding values in Komeili et al. [22], 95 and 94%, respectively (Fig. 5). The reduction in the overall accuracy (i.e. 82 to 65%) in identifying the severity of the curve using ST parameters in this study with respect to Komeili et al. [22] was caused by a larger false positive error rather than false negative error, i.e. classifying a considerable number of the deviation patches of patients with mild scoliotic curves in the moderate/severe group. One third of the misclassified mild curves had a Cobb angle in the range of 20–25 degrees, which is on the margin of the mild-moderate severity definition used in the scoliosis clinic. It should be noted that, there is a ±5° inter- and intra-observer error in the measurements of Cobb angle [23, 24], therefore there is a possibility that some of the misclassified mild curves with Cobb angle in the range of 20–25 degrees could have been diagnosed as moderate if the radiograph assessment had been repeated.

The overall sensitivity of 95% in Fig. 5 indicates a high ability to detect moderate/severe curves which are more susceptible to progress than mild curves and thus do require further attention. Taking full torso radiograph in the follow-up visits is the standard for all patients at the scoliosis clinics, hence, subjects whose classification was false positive would not get more radiographs than others without ST assessment, and we would still prevent radiograph exposure for some truly negative cases. Komeili et al. [15] reported a predicted 44% reduction in the radiograph acquisition if they combined the ST analysis with the radiograph assessment in the monitoring of curve progression. Because of the slightly lower accuracy observed in this study in screening for curve progression, we estimate a 5% lower reduction of the need for radiograph compared with the work of Komeili et al. [15]. Nevertheless, the asymmetry analysis improvements suggested in this study would result to a more conservative monitoring. Because of the higher sensitivity obtained (Fig. 5b) a lower number of patients would suffer from missing the opportunity of an early diagnosis of curve progression by using our improved ST analysis.

We had enough data for statistical analyses, however we have to acknowledge that only 39 patients in the Training group had curve progression, which may not represent the diversity of scoliotic curve types, such as single and double curves. Another limitation of our study is exclusion of trunk axial rotation in our predictions, which is an important piece of information for clinicians in prescribing patient specific braces. It may be possible to improve the accuracy of curve progression diagnosis if more ST parameters, such as curvature of back valley [25] and trunk rotation [26], were combined with the RMS and MaxDev parameters included in this study. The strong correlation between deviation patches and spinal curve location can also provide good information about the kyphosis and lordosis angles obtained from radiographs, which, however, may not be necessary if the proposed method is used only in scoliosis cases clinics. Our asymmetry analysis is compatible with the ST database of those clinics that do not use the same acquisition system in capturing the full torso geometry as we do. The markerless feature of the asymmetry analysis reduced the common limitations of ST methods, such as dependency of the method to specific local marker-based measurements, and data acquisition technique. Having the 3D geometry of the torso, in normal posture, is the only required condition for using this method. Our strategy in classifying curve severity and curve progression could also be followed in other ST classification systems, in which the cosmetic parameters of the torso are correlated to the scoliotic curve characteristics. Our conservative approach in minimizing the number of false progression and false moderate/severe in the classification system may result in a lower overall accuracy, however it would reduce the risk of missing a progressed or moderate/severe curve, which is the main concern in replacing radiographs with ST assessments in scoliosis clinics.

The modified thresholds used to define asymmetry patches successfully separated the deviation patches in the upper two-third (T-TL) section from the lower one-third part (Lumbar) section and eliminated the manual work, which was previously necessary for isolating the deviation patches in some patients. The modified thresholds used in this study, allow automation of the analysis and led to the same level of sensitivity in identifying the curve severity as in our previous work. Similarly, the modified analysis led to similar sensitivity for monitoring the progression of the scoliotic curve as our previous work. Despite the novel method being unfortunately associated with a higher risk of misclassification of cases with no progression, significant numbers of patients (approximately 39%) may be able to avoid radiographs confident that their curves did not progress.