Vertebral Wedging Index (VWI) in AIS

• Vertebral Wedging Index (VWI) is a morphology-based metric that quantifies intra-vertebral wedging in AIS through predicted endplate angles.
• It is derived from a deep learning pipeline integrating HRNet, Swin-Transformer modules, and SVD-based curve detection to accurately identify vertebral deformation.
• VWI serves as a complementary biomarker to the Cobb angle by providing localized prognostic insights for early detection and treatment planning in AIS.

The Vertebral Wedging Index (VWI) is a morphology-based metric introduced for adolescent idiopathic scoliosis (AIS) to quantify intra-vertebral wedging across the vertebrae that belong to a detected scoliotic curve. In the framework presented in "Accurate Cobb Angle Estimation via SVD-Based Curve Detection and Vertebral Wedging Quantification," VWI is derived from predicted superior and inferior endplate angles rather than predicted directly, and is positioned as a structural complement to the Cobb angle, which remains the gold standard for overall coronal curve magnitude assessment. The same study situates VWI within an automated AIS pipeline built on an HRNet backbone, Swin-Transformer modules, biomechanically informed constraints, and SVD-based curve detection, evaluated on 630 full-spine anteroposterior radiographs from patients aged 10–18 years; AIS is described there as affecting approximately 2.2% of boys and 4.8% of girls worldwide (Shi et al., 29 Sep 2025).

1. Definition and mathematical formulation

VWI is defined as the average absolute difference between the superior and inferior endplate angles of the vertebrae within a detected curve. The paper gives the metric as

$$\text{VWI} = \frac{1}{N}\sum_{i=s}^{s+N-1}\left|\theta_i^{\text{upper}} - \theta_i^{\text{lower}}\right|$$

where $s$ is the superior end vertebra index, $N$ is the number of vertebrae in the curve, $\theta_i^{\text{upper}}$ is the predicted angle of the superior endplate of vertebra $i$, and $\theta_i^{\text{lower}}$ is the predicted angle of the inferior endplate of vertebra $i$ (Shi et al., 29 Sep 2025).

The intended geometric interpretation is direct: when a vertebral body is more wedge-shaped, its superior and inferior endplates are less parallel, and the absolute angular difference increases. VWI therefore quantifies vertebral body wedging rather than global spinal curvature. In the paper’s framing, this distinction is central: the metric is explicitly constructed to capture vertebral deformation across a curve segment, not merely the aggregate angular extent of the deformity.

Because the averaging is performed over vertebrae indexed from $i=s$ to $i=s+N-1$, VWI is curve-specific rather than spine-global. Its value depends on identifying which vertebrae belong to the relevant curve and on estimating the superior end vertebra index and the curve length. This makes VWI a local morphologic descriptor tied to a curve segment rather than a single summary statistic for the entire radiograph.

2. Derivation from vertebral morphology predictions

The network does not predict VWI as a standalone output. Instead, it predicts two classes of vertebral measurements for each vertebra: upper and lower endplate angles, and midpoint coordinates of those endplates. The paper formalizes the mapping as

$$f_{\Theta}: \mathbf{I} \mapsto \left\{ \left(\theta_i^{\text{upper}}, \theta_i^{\text{lower}}, \mathbf{p}_i^{\text{upper}}, \mathbf{p}_i^{\text{lower}}\right)\right\}_{i=1}^N$$

with $s$0 the spinal radiograph, and $s$1 and $s$2 the midpoint coordinates of the endplates (Shi et al., 29 Sep 2025).

This output design is motivated by the stated aim of preserving the anatomical reality of vertebral wedging in progressive AIS. The output head jointly predicts heatmaps for landmark localization and vector fields for endplate angle regression. VWI is then computed from the predicted endplate angles after the vertebrae belonging to a detected curve have been identified.

Within the broader system, these predictions are produced by a deep learning framework that combines an HRNet backbone with Swin-Transformer modules and biomechanically informed constraints for enhanced feature extraction. In the reported evaluation on 630 full-spine anteroposterior radiographs with rigorous dual-rater annotation, the method achieved 83.45% diagnostic accuracy and 2.55° mean absolute error, while also demonstrating exceptional generalization capability on out-of-distribution cases (Shi et al., 29 Sep 2025).

3. Role within the SVD-based curve detection framework

Although VWI itself is computed from same-vertebra angle differences, it is embedded in a larger representation based on pairwise angular relationships. The paper constructs an angle matrix

$s$3

where $s$4 and $s$5 are row vectors of upper and lower endplate angles, respectively, and $s$6 is an 18-dimensional all-ones vector. Each element is

$s$7

This matrix encodes pairwise angular relationships between endplates across vertebrae (Shi et al., 29 Sep 2025).

The SVD-based pipeline proceeds in five stages: predict upper and lower endplate angles and midpoints for each vertebra; build the angle matrix $s$8; apply Singular Value Decomposition, $s$9; use the first principal component $N$0 to identify local extrema; and treat local maxima and minima as candidate end vertebrae of spinal curves. The key methodological claim is that the framework does not assume a fixed number of curves or a predefined curve pattern; rather, it infers curve structure from the intrinsic organization of the predicted endplate-angle matrix.

VWI enters only after this curve detection stage. Once a set of vertebrae has been assigned to a curve, the same vertebral morphology representation used for curve identification is reduced to the within-vertebra absolute angular differences that define the index. This coupling is important: VWI depends operationally on the same endplate prediction pipeline that underlies automatic curve segmentation, yet it is not itself an SVD score or a principal-component statistic.

4. Distinction from Cobb angle and structural interpretation

The paper is explicit that Cobb angle and VWI quantify different aspects of scoliosis. Cobb angle measures the overall coronal curvature magnitude of the spine by using the most tilted end vertebrae, whereas VWI measures vertebral body deformation, specifically the degree of wedging at each vertebral level within the curve (Shi et al., 29 Sep 2025).

Aspect	Cobb angle	VWI
Measurement target	Overall coronal curvature magnitude	Vertebral body deformation
Structural level	Curve-level metric	Morphology-level metric
Construction	Most tilted end vertebrae	Average absolute upper–lower endplate angle difference

This distinction is clinically significant in the study’s interpretation of AIS progression. The paper notes that progression involves early intervertebral disc changes, followed by structural modifications to vertebral bodies producing wedge-shaped deformities. On that account, VWI is intended to capture a more proximal structural sign of progression than Cobb angle. The article’s examples include cases with similar Cobb angles but different VWI values, indicating that patients with comparable curve magnitudes may nonetheless differ substantially in vertebral deformation burden.

A common misconception addressed by the paper is that a sufficiently informative curve magnitude metric should subsume morphology. The proposed use of VWI rejects that equivalence. The authors present VWI not as a replacement for Cobb angle, but as a complementary metric that adds anatomical specificity and prognostic insight. This suggests a dual-description paradigm in which one measure indexes global coronal deformity and the other indexes the local structural remodeling associated with progression.

5. Longitudinal evidence and prognostic significance

The paper reports a longitudinal analysis of 138 patients with multiple follow-up examinations and a mean follow-up of $N$1 months. Baseline VWI was compared with subsequent Cobb angle progression, and the reported associations favored VWI over baseline Cobb angle as a prognostic correlate (Shi et al., 29 Sep 2025).

The correlations are given as follows:

$N$2

$N$3

$N$4

In the study’s interpretation, initial VWI showed significant correlation with curve progression, initial Cobb angle showed no significant correlation, and initial Risser score was significant but weaker than VWI. The authors treat this as evidence that VWI has prognostic value in their cohort.

The paper further argues that this prognostic behavior supports several use cases: early AIS detection, treatment planning, and monitoring progression. A specific clinical example is a complex congenital scoliosis case in which a moderate Cobb angle would not ordinarily suggest surgery, but the patient had a high VWI of 10.1°, and expert assessment supported operative treatment. Within the study’s logic, this example illustrates how Cobb angle alone may miss clinically important structural severity. A plausible implication is that morphology-sensitive biomarkers such as VWI may be particularly useful in cases where curve magnitude underrepresents deformity burden.

6. Scope, assumptions, and limitations

Several implementation constraints delimit the interpretation of VWI. First, the metric depends on accurate endplate-angle prediction: because it is computed from $N$5 and $N$6, any error in vertebral landmark localization or angle regression propagates directly into the index. Second, VWI is curve-specific; it is computed only over vertebrae belonging to a detected curve, so end vertebra identification is a prerequisite rather than an optional postprocessing step (Shi et al., 29 Sep 2025).

Third, the vertebral wedging interpretation is an inference from radiographic geometry. The paper explicitly states that the metric is based on the idea that the angle between a vertebra’s superior and inferior endplates reflects wedging. This is described as clinically sensible in progressive AIS, but it remains a radiographic surrogate rather than a direct volumetric measure of vertebral shape. Fourth, the measure is derived from 2D standing PA radiographs, not 3D imaging. VWI is therefore a 2D radiographic surrogate for vertebral morphology.

Fifth, the reported implementation is tied to a particular annotation scheme. The model was trained on 18 vertebrae, from C7 to L5, with four landmarks per vertebra. As a consequence, the indexing used in VWI computation is not annotation-agnostic. Sixth, unlike the curve detection stage, which uses a 10° clinical significance threshold for $N$7, VWI itself is not thresholded in the paper. It is a continuous severity measure. This matters methodologically because VWI is not presented as a binary decision rule; its interpretation depends on clinical context, longitudinal follow-up, and its combination with curve-level assessment.

Taken together, these constraints place VWI in a specific epistemic category: a curve-local, model-derived, 2D morphologic index whose validity depends on the fidelity of endplate prediction and curve delineation. Within that scope, the paper presents it as a promising complementary biomarker for early detection, prognostic assessment, individualized treatment planning, and monitoring of AIS progression (Shi et al., 29 Sep 2025).