Introduction

Variations in the two-dimensional (2D) and three-dimensional (3D) spinal curve patterns in adolescent idiopathic scoliosis (AIS) have been used to classify, guide clinical management of the patients, and predict treatment outcomes [1,2,3,4]. Several studies have determined surgical guidelines based on the King or Lenke classification [1, 2, 5,6,7,8,9,10,11,12,13,14]. Anecdotally adverse surgical outcomes have been observed following these guidelines [13, 15,16,17]. A classification system that can differentiate between the key 3D curve characteristics and their association with the surgical outcomes is of critical need.

Lenke classification is currently the most common classification system for AIS [1]. This system uses the location (proximal thoracic, thoracic, or lumbar), number, and flexibility of the 2D frontal curves and adds sagittal and lumbar modifiers to determine the curve types [1]. Other studies considered the 3D parameters of the spine, particularly the axial rotation of the curves, and determined several subtypes within each Lenke type [18,19,20]. Classifications of the AIS based on the 3D vertebral positions have been developed [4, 18, 20,21,22]; however, the 3D alignment of the vertebral bodies in frontal, sagittal, and axial planes has not been included in the AIS classification. As the patient-specific vertebral relative alignments and disc wedging guide the surgical moves, it remains to determine if considering the patterns of 3D alignments of the vertebral bodies at pre-operative, in addition to their 3D positions, can lead to a classification of AIS that is deterministic of the surgical outcomes. Moreover, the surgical outcome evaluation of AIS mainly focuses on one outcome at the time, while methods to assess the global alignment of the spine are not fully explored [9, 13, 21, 23].

To this end, we aimed to develop a classification system of the 3D spinal alignment before and at two-year follow-up of the posterior spinal fusion (PSF) using the 3D vertebral alignment and position. We further determined the distribution of the 3D pre-operative clusters in the 3D outcome groups. We hypothesized that a pre-operative classification of the spinal curve based on the 3D position and alignment of the vertebral bodies relates to the 3D global alignment of the spine at 2-year follow-up at a higher accuracy rate compared to the Lenke classification.

Methods

Patient population

AIS patients who had received a PSF with at least 2-year follow-up were reviewed in a multicenter (three) and multi-surgeon (seven) data registry. Patients with two-view (posterior–anterior and lateral) spinal stereoradiographs at pre-operative and 2-year follow-ups were included. Lenke type classification was extracted from the database. A total of 20 non-scoliotic adolescents were included.

Image processing and variable extraction

3D reconstructions of the spine and pelvis radiographs were generated in SterEOS 2D/3D (EOS imaging, Paris, France) for pre- and post-operative images. Using the vertebral endplate alignments [24], the 3D model of the spine was used to measure the T1–T12 and T5–T12 kyphosis, L1–S1 lordosis, pelvic incidence (PI), proximal and main thoracic and lumbar coronal Cobb angles, and pelvic incidence. T1 to L5 vertebral 3D alignments, i.e., frontal tilt (ΘFrontal), sagittal tilt (ΘSagittal), and axial rotation (ΘRotation), were calculated using the average of the vectors connecting the posterior and anterior points on the vertebral endplate on the superior and inferior endplates (Fig. 1) [25]. The angle between this average vector and the true X-, Y-, Z-axes determined the frontal, sagittal, and axial alignment of the vertebrae (Fig. 1). The vertebral centroids were calculated using the 3D reconstruction models [25]. The image analysis extracted a total of 6 variables per vertebra, 102 variables per spine (Fig. 1). The X-, Y-, Z-coordinates were scaled isotropically in a way to achieve normalized height [4, 20]. The vertebral 3D alignments were standardized (divided by standard deviation and the mean was removed) to eliminate the effect of the severity of the curve, which allowed performing the classification based on the deformity patterns, solely.

Fig. 1
figure 1

Vertebral 3D position (X, Y, Z) and 3D alignment parameters (Frontal and sagittal tilt and axial rotation). A vector constituted of these 6 variables was generated per vertebral level (17 vertebrae T1–L5). These variables were used in the classification algorithm. The origin of the coordinate system is the posterior aspect of the sacral plateau

Full size image

Clustering analysis

The 102 variables per patient were used to determine the pre- and postoperative 3D clusters. A fuzzy C-means clustering with 1000 non-repetitive start points was used to calculate the degree of membership of patients to each cluster. The maximum degree of membership determined the cluster number. The cluster analysis was performed for pre- and post-operative spines separately.

Each cluster was described based on the curve pattern characteristics in three anatomical planes. Hypo (< 10°), hyper- (> 40°), and normal kyphosis (between 10° and 40°) were defined according to Lenke classification [1]. In each cluster, the axial projection was presented based on a previous classification of the curves projection on the transverse plane [20, 26]: in the loop shaped axial projection, the vertebral rotation outside of the main curve does not change; thus, the vertebrae are rotated in the same direction or not rotated at all. In the lemniscate type, the direction of the vertebral rotation is different between the main and compensatory curves resulting in thoracic and lumbar vertebrae rotated into opposite directions projecting a lemniscate shape on the axial plane.

Statistical analysis

The spinal and pelvic measurements in each of the pre- and post-operative clusters were statistically compared. The distributions of the pre-operative clusters in the Lenke types and in the 3D outcome clusters were determined.

Results

A total of 371 AIS patients were included. All Lenke types were included: Lenke 1 (35%), Lenke 2 (23%), Lenke 3 (10%), Lenke 4 (7%), Lenke 5 (15%), and Lenke 6 (10%). The clustering analysis determined nine pre-operative 3D clusters.

Pre-operative clusters

Table 1 summarizes the spinal and pelvic measurements in the pre-operative clusters. Analysis of variance for each variable showed significant differences between the clusters (Table 1). Figure 2 shows the frontal, lateral, and axial views of the preoperative clusters.

Table 1 The spinal and pelvic measurements in the preoperative subtypes
Full size table
Fig. 2
figure 2

Pre-operative clusters of the AIS cohort. A total of nine 3D curve patterns were identified. a Types 1–4: right thoracic curves. b Types 5–7 left thoracolumbar/lumbar curves. c Type 8: right low apex thoracic/thoracolumbar and d Type 9: left thoracic right lumbar curve. Summary of the spinal parameters for each cluster is listed in Table 1. The average spinal alignment in the cohort of 20 non-scoliotic adolescents is included

Full size image

The nine pre-operative group description in 3D, frontal plane, sagittal plane, and axial plane are as follows:

(A) Patients appear with a right thoracic curve, which exceeds the left lumbar Cobb angle in frontal plane; includes Type 1 to Type 4 curve patterns. 65% of the patients belonged in this group.

Type 1—two 3D curves. Hyperkyphotic. The axial projection is lemniscate shaped.

Type 2—one 3D curve. Kyphosis was in the hypokyphotic–normal range, long frontal thoracic curve with small or no rotation in lumbar spine. Highest pelvic incidence among the subtypes. The axial projection is loop shaped.

Type 3—two 3D curves. Thoracic and lumbar cures are close in magnitude. Normal kyphosis. The axial projection is lemniscate shaped.

Type 4—one 3D curve. Hypokyphotic with long lordotic curve. The axial projection is loop shaped.

(B) Patients with a left lumbar curve, which exceeds the right thoracic Cobb angle; includes Type 5 to Type 7 curve patterns. 27% of the patients belonged in this group.

Type 5—one 3D curves. Hyper to Normal kyphosis. Long kyphotic curve. The axial projection is loop shaped.

Type 6—two–three 3D curves. Normal to Hypokyphosis. The axial projection is lemniscate shaped.

Type 7—two–three 3D curves. Hypokyphotic. Flat T4–T12 kyphosis. Lowest pelvic incidence. The axial projection is lemniscate shaped.

(C) Type 8—low apex right thoracic curve. One 3D curves. The apex of the thoracic curve is lower than the subtypes in section A (right thoracic curves). Normal–hyper kyphosis. The axial projection is loop shaped.

(D) Type 9—left thoracic/right lumbar. One 3D curve—Hyperlordotic. Long lordotic curve. The axial projection is loop shaped.

The distribution of the Lenke types in each of the 3D pre-operative clusters is shown in Table 2. These most common Lenke types comprised at least 25% of the patients in each of 3D clusters (Table 2).

Table 2 Distribution of the nine preoperative 3D clusters based on the Lenke criteria
Full size table

Post-operative clusters

Three post-operative 3D clusters were determined based on our clustering analysis with a distribution of Type 1 (41%), Type 2 (35%), and Type 3 (24%) (Fig. 3). The average frontal, sagittal, and axial spinal and pelvic parameters in each outcome subtypes are presented in Table 3. Type 1 and type 2 had similar frontal profile while Types 1 and 3 had similar sagittal profile (Fig. 3 and Table 3). Type 1 had the highest proximal thoracic Cobb angle and lowest PI and Type 3 had the lowest proximal thoracic Cobb and the highest lumbar Cobb angle and rotation, p < 0.05 (Table 3). Type 2 had the lowest T5–T12 kyphosis and pelvic incidence–lordosis mismatch (PI-LL) compared to the other two types, p < 0.05 (Table 3).

Fig. 3
figure 3

Post-operative clusters of the cohort. A total of three 3D curve patterns were identified. Residual proximal thoracic in Type 1, hypokyphosis in Type 2, and residual lumbar Cobb angle and rotation in Type 3 were the main characteristics. Summary of the spinal parameters for each cluster are listed in Table 3

Full size image
Table 3 The spinal and pelvic measurements in the subtypes at 2-year follow-up
Full size table

Link between the pre-operative cluster (3D and Lenke) and post-operative 3D clusters

A majority (50% and higher) of the patients in the pre-operative 3D types 1, 3, 4, 8, and 9 had the Type 1 outcome profiles (Table 4). 68% of pre-operative cluster Type 2 had the Type 2 outcomes. More than 50% in Types 5, 6, 7 had type 3 outcomes (Table 4).

Table 4 Proportion of the 3D preoperative clusters in the three outcome clusters
Full size table

Considering the distribution of the Lenke types in the outcome groups, more than 50% of the Lenke Types 2, 5 and 6 ended up in one outcome group (Table 5) while the relationship between Lenke 1, 3, and 4 and the outcome clusters was inconclusive (Table 5).

Table 5 Proportion of the Lenke types in the three outcome clusters
Full size table

Discussion

A 3D classification system that distinguishes between the pre-operative curve patterns and associates pre-operative curve types with the post-operative global spinal alignment is the first step in developing data-driven models for PSF outcome prediction. We developed a classification system based on the 3D curve patterns of the AIS patients before and two years after surgery. Considering these patterns, we discovered strong associations between the spinal alignment at pre-operative and 2-year follow-ups.

Classifications of the spinal deformity in AIS are mainly based on 2D radiographs [1, 2]. Lenke classification uses 6 frontal curve patterns, three kyphosis modifiers, and three lumbar modifiers [1]. Eliminating the non-plausible combinations (Lenke 5A and 6A for example), approximately 45 different combinations of the Lenke types with modifiers can be determined. Our analysis showed a total of nine 3D subtypes, each composed of several Lenke types (Fig. 2; Table 2). While we determined the dominant Lenke type within each of these 3D clusters, the percent of patients with the same Lenke type in the 3D clusters remained low and exceeded 50% in only two subtypes (Table 2). Our classification specially draws attention to variation in the axial and sagittal curvature of the spine in patients with thoracic or thoracolumbar curves (Fig. 2). The axial patterns determined the number of 3D curves in loop versus lemniscate shaped axial projection types (Fig. 2). The sagittal curves of the subtypes showed a high variation in the position of the inflection point, a factor that was not accessible in the Lenke classification which can better guide the selection of the fusion levels and imparting kyphosis/lordosis. Our results showed considering the 3D clusters can determine different curve patterns within the Lenke types, which in turn can facilitate communicating these differences between the surgeons. It should be kept in mind that variations between the individuals, scoliotic and non-scoliotic alike, within a cluster is expected (Fig. 2).

The role of the 3D descriptors of the scoliotic curves, particularly the axial rotations, has been underlined in the literature [18,19,20, 24, 27]. Despite the increased use of stereoradiography in orthopedic clinics, the application of the 3D models of the spine in clinical diagnostics remains limited. Vertebral rotation [18], orientation of the plane of maximum curvature [18, 28], global torsion [22] and writhe [29] of the spinal centerline were used to identified subtypes of AIS patients. All these methods, although determined fundamental differences between the curve patterns within the AIS types, required intensive image post processing and analysis[29]. Our classification uses the 3D position and alignment of the vertebral bodies to determine the 3D clusters, but subsequently describes the curve patterns in the frontal, sagittal, axial characteristics (loop, lemniscate) using 2D radiographs of the spine (Fig. 2). A classification system that relies on a few curve patterns reduces the need to excessive image processing and can facilitate the clinical application of this classification.

In the current analysis, we determined two distinct axial projection patterns of the curves in our cohort of all Lenke types (Fig. 2). These patterns, i.e., loop versus lemniscate axial projections were also observed in previous classifications of the scoliosis [16, 20, 26] and determined the number of true 3D curves [20, 26]. In group A (Types 1–4) and group B (Types 5–7) two 3D curves was observed in Types 1, 3, 6, 7 (lemniscate axial projection), while in the types with loop shaped axial projection (Types 2, 4, 5) the vertebral rotation in the secondary curve are insignificant showing only one 3D curve. It was shown before that a correct identification of these 3D curves extents can impact the outcomes of the PSF [16]. It was shown when a 3D curve was included in the fusion While the measurements of the vertebral axial rotation on 2D radiographs is less reliable [30], the patterns of changes in the pedicle orientations can be detected with high accuracy (kappa = 0.98) using only the 2D radiographs and can be used to determine the AIS subtypes [31]. The low apex thoracic curve patients (Type 8) show essentially different curve patterns when compared with groups A and B clusters, as was also suggested before [32]. Our analysis separated this cohort (Type 8) from thoracic curve groups (Types 1–4) showing that although Type 8 has a similar axial pattern to Types 2 and 4 but the sagittal and frontal curves do not match any of these curve patterns revealing differences between the low apex thoracic and typical thoracic and thoracolumbar curves (Fig. 2). Type 9 not only had a left thoracic curve but also the sagittal and axial curves were different from the other thoracic types suggesting a significantly different 3D curve pattern (Fig. 2).

Instead of focusing on one radiographic outcome, we clustered the 3D spinal alignment at 2-year follow-ups to determine significantly different global spinal alignments as our surgical outcome measure (Fig. 3). While the pre-operative 3D clusters related to the outcome curve patterns (Table 4), the role of the surgical factors in achieving specific outcomes cannot be ignored [21]. In the current cohort, while variation in the surgical technique existed between the surgeons, a majority of the patients received the same treatment [33]. These surgical moves/techniques included: intraoperative traction, rod tensile strength (200 KSI both rods), all screws construct type, in situ bending, Ponte Osteotomy, wide posterior release, segmental derotation with two rods in, and both ends locked before direct vertebral rotation. Identifying the surgical parameters, specific to each of the 3D pre-operative clusters, can increase the likelihood of achieving one of the outcome types. This is the subject of our future study.

A previous 3D classification of all Lenke types used 2D measurements of the spine and the plane of maximum curvature as variables and determined 11 subtypes of scoliosis [28]. The subtypes were characterized as low, average, and high for frontal curves and hypo, hyper, neutral for sagittal curves. As a result, this classification determined subtypes by relying on curve magnitudes [28]. Our classification, however, used the patterns of the normalized values of the vertebral position and alignment thus eliminating the effect of the curve severity in our classification while allowed identifying different curve patterns (Table 1; Figs. 2 and 3). As seen in Table 1, the variations in the pre-operative frontal curve magnitudes in our classification were smaller than the previous 3D classification system [28].

The relationship between the pre- and post-operative clusters remains to be tested for a verity of surgical techniques to better evaluate reproducibility and reliability of the proposed classification system. External validations of the method to verify if these curve patterns can be learned and used by orthopedic surgeons are warranted.

Conclusion

A new AIS classification of the pre- and post-operative spinal 3D curve patterns was developed. We discovered a relationship between the pre-operative clusters and the post-operative groups, showing the clinical relevance of this classification and a promising application of this classification in PSF outcome prediction.