Fully Automated Deep Learning-Based Lenke Classification for Adolescent Idiopathic Scoliosis Using Multi-View Full-Spine Radiographs: Development and Clinical Validation

Abstract

Study Design

Retrospective validation study.

Objectives

To develop and validate a fully automated deep learning framework for radiographic measurement and Lenke classification in adolescent idiopathic scoliosis (AIS) using multi-view full-spine radiographs.

Methods

Consecutive patients with AIS from 3 hospitals who underwent standardized 4-view full-spine radiography between January 2019 and January 2026 were retrospectively reviewed. The automated framework incorporated vertebral detection, vertebra-level landmark estimation, radiographic parameter computation, and rule-based Lenke classification. Expert manual measurements and consensus Lenke classification served as the reference standard. Performance was assessed on an independent test set using detection and keypoint metrics, Cobb angle measurement agreement, classification accuracy, clinician agreement, and workflow efficiency.

Results

In 76 independent test cases, the framework achieved a vertebral detection AP@0.75 of 0.858 and an overall keypoint AP of 0.951. Cobb angle measurement showed a mean absolute error of 2.4°, with excellent agreement with expert measurements (ICC = 0.991; Pearson correlation = 0.987; Spearman correlation = 0.983). Accuracy was 0.895 for Lenke curve type, 0.868 for lumbar modifier, 0.908 for sagittal thoracic modifier, and 0.816 for the complete Lenke label. Agreement between the AI system and the senior surgeon was high ( κ = 0.873 for curve type; κ = 0.781 for complete label). Mean processing time was 12.3 seconds per case, and AI-assisted review reduced mean review time from 8.2 to 1.5 minutes per case.

Conclusions

This fully automated deep learning framework achieved accurate radiographic measurement and Lenke classification with strong agreement with spine surgeons and marked efficiency gains. The method may serve as an interpretable decision-support tool for preoperative AIS assessment.

Introduction

Adolescent idiopathic scoliosis (AIS) is the most common spinal deformity in adolescents and is characterized by a three-dimensional deformity involving the coronal, sagittal, and axial planes.^1,2 Recent epidemiological studies and meta-analyses suggest that scoliosis in children and adolescents is common at the population level, with prevalence estimates around 3%, highlighting its public health relevance. Beyond radiographic deformity, AIS may be associated with pain, impaired appearance perception, psychosocial distress, reduced health-related quality of life, and treatment-related burden during a critical developmental period.^3,4

Accurate radiographic assessment is fundamental to treatment planning, particularly for patients being considered for operative correction. ⁵ Conventional full-spine radiographs remain the cornerstone of preoperative evaluation because they provide key information regarding curve magnitude, global alignment, and flexibility. ⁶ Among available classification systems, the Lenke classification is one of the most widely used frameworks for preoperative assessment and surgical decision-making in AIS. ⁷ Originally developed to guide the extent of spinal arthrodesis, the Lenke system integrates coronal curve pattern, lumbar modifier, and sagittal thoracic modifier into a unified surgical classification and has become a practical template for fusion-level selection. ⁸ By incorporating standing coronal, lateral, and side-bending radiographs, it provides a more comprehensive description of deformity than Cobb angle measurement alone.

Despite its clinical utility, manual Lenke classification is complex and time-consuming. Surgeons must review multiple radiographic views, identify end vertebrae, measure Cobb angles across several curves, determine structural versus nonstructural components on bending films, and assign lumbar and sagittal modifiers according to predefined thresholds. This process depends heavily on reader experience. Prior studies have shown that scoliosis classification systems, including Lenke-based assessment, are affected by interobserver and intraobserver variability.^9,10 Differences in image interpretation and diagnosis may also exist between spine surgeons at top-tier clinical centers and those at secondary medical centers. ¹¹ In addition, Cobb angle determination itself remains vulnerable to inconsistency, and small threshold-based differences may lead to different classification decisions.^12,13

Recent advances in artificial intelligence (AI) and deep learning have shown promise in scoliosis image analysis. Contemporary studies have reported automated or semiautomated approaches for Cobb angle measurement, sagittal parameter measurement, and related radiographic assessment tasks in AIS. ¹⁴ However, most published studies have focused on isolated subtasks rather than the integrated multi-view reasoning required for operative Lenke classification. A clinically useful automated Lenke framework must integrate multi-view full-spine radiographs, extract anatomically meaningful landmarks, perform rule-based radiographic measurements, and generate a final classification that remains interpretable to spine surgeons. ¹⁵ More recently, investigators have begun to address automated Lenke classification directly using deep learning, and multi-view semantic segmentation and other deep learning pipelines have demonstrated the feasibility of automated preoperative Lenke classification in AIS. ¹⁶ Nevertheless, the number of such studies remains limited, and additional validation using technically distinct and clinically interpretable workflows is still needed. ¹⁷ In particular, a landmarkbased framework may offer practical advantages because it better mirrors the conventional radiographic reasoning process used by surgeons, thereby enhancing transparency, facilitating error analysis, and supporting integration into routine preoperative assessment.^14,16

Therefore, the purpose of the present study was to develop and clinically validate a fully automated deep learning-based framework for preoperative radiographic measurement and Lenke classification in AIS using multiview full-spine radiographs, including standing coronal, lateral, and side-bending views. We further compared the automated results with expert manual assessment to evaluate measurement agreement, classification performance, and the potential clinical applicability of this approach in a standardized preoperative workflow.

Materials and Methods

This retrospective study was conducted using radiographic data collected from Beijing Anzhen Hospital, Peking University First Hospital, and Beijing Chaoyang Hospital. Consecutive patients with adolescent idiopathic scoliosis who underwent standardized four-view full-spine radiographic examination between January 2019 and January 2026 were screened for eligibility. The study was approved by the institutional ethics committee of Beijing Anzhen Hospital, and the requirement for informed consent was waived because of the retrospective use of de-identified data.

Inclusion and Exclusion Criteria

From the total 507 screened patients from three hospitals (Beijing Anzhen Hospital, 90; Peking University First Hospital, 225; Beijing Chaoyang Hospital, 192), eligible patients were randomly partitioned at the patient level into training, validation, and independent test sets using stratified sampling by center and Lenke type. The split ratio was approximately 70%/15%/15%, resulting in 355, 76, and 76 patients, respectively. To prevent data leakage, all four-view radiographs belonging to the same patient were kept together in a single subset. The test set remained completely untouched during model development and was used only for the final evaluation. Patients were included if they met the following criteria: (1) diagnosis of adolescent idiopathic scoliosis; (1) availability of complete standing posteroanterior, lateral, left-bending, and right-bending fullspine radiographs; and (3) radiographic image quality adequate for measurement and classification. Patients were excluded if they had non-idiopathic scoliosis, incomplete radiographic examinations, severe imaging artifacts, or unavailable expert reference labels.

Deep Learning-Based Automated Classification Methods

We developed a fully automated deep learning framework for Lenke classification in adolescent idiopathic scoliosis (AIS) using four-view full-spine radiographs. The framework was designed to process anteroposterior, lateral, left-bending, and right-bending radiographs and to generate both quantitative radiographic measurements and final Lenke labels without manual intervention. The overall workflow consisted of image preprocessing, vertebral localization, vertebra-level landmark estimation, radiographic parameter computation, and rule-based Lenke classification.^18,19

As shown in Figure 1, all radiographs first underwent standardized preprocessing, including DICOM anonymization, grayscale normalization with window clipping and z- score standardization, orientation correction, spatial standardization to 1024 × 2048 pixels, and contrast enhancement within the spinal region. The preprocessed images were then analyzed by a vertebral detection model, termed SpineDetector, which was built on an HRNet-W32 backbone with a CenterNet detection head.^20,21 This model was used to localize the center positions and boundary regions of individual vertebrae on full-length spine radiographs. For severe scoliosis cases with Cobb angles greater than 60°, a multiscale inference strategy was introduced during detection. Specifically, inference was performed at both the original resolution and a 1.5 × magnified resolution, and the resulting predictions were merged using non-maximum suppression to improve vertebral detection in regions with marked overlap.

Figure 1.

Framework of the proposed automated Lenke classification for adolescent idiopathic scoliosis. AP: anteroposterior radiograph; LAT: lateral radiograph; LB: left-bending radiograph; RB: right-bending radiograph

The localized vertebral regions were subsequently cropped and processed by a vertebra-level keypoint detection model, termed VertebraKeypointNet, based on a stacked hourglass architecture. ²² This model estimated vertebral corner points, medial pedicle landmarks, and the midpoint of the superior S1 endplate. To reduce CSVL-related classification errors, an additional S1-specific prediction branch was incorporated into the optimized keypoint model by extending the original heatmap representation with a dedicated channel for direct localization of the superior S1 midpoint. This design enabled more accurate geometric characterization of both coronal and sagittal spinal alignment. ¹⁹

Based on the predicted vertebral and sacral landmarks, the system automatically computed the radiographic parameters required for Lenke classification, including the Cobb angles of the proximal thoracic, main thoracic, and thoracolumbar/lumbar curves, the T5–T12 kyphosis angle, and the positional relationship between the central sacral vertical line (CSVL) and the lumbar apical vertebra.^19,23-25 The structurality of each curve was determined according to curve magnitude and the degree of correction on bending radiographs. Based on these computed parameters, the final Lenke classification was generated automatically by a rulebased decision module. The major curve was identified according to the largest coronal Cobb angle, whereas the remaining curves were evaluated according to their magnitude and structurality. The lumbar modifier was assigned according to the CSVL-related coronal alignment, and the sagittal thoracic modifier was determined from the T5–T12 kyphosis. The system then generated the final Lenke type, lumbar modifier, and sagittal thoracic modifier for each case. ²⁵

For cases near clinically relevant decision thresholds, the system automatically generated a manual review recommended flag to indicate potential classification uncertainty. In the optimized system, this flag was assigned to cases with Cobb-angle measurements within ± 3° of the structural/non-structural or sagittal-modifier thresholds.

To facilitate clinical use, the two optimized models were further integrated into a browser/server architecture with a Python Flask-based backend, a PyTorch inference engine, ²⁶ asynchronous task scheduling via Celery and Redis, and structured output generation. Following model inference and parameter computation, the system automatically generated a structured JSON report, annotated overlay images, and a human-readable technical summary. The complete four-view analysis workflow, including report generation, required a mean processing time of 12.3 seconds per case. The deployed workflow also supported PACS integration through automated DICOM reception, viewtype recognition, case assembly, result write-back, and notification of cases flagged for manual review.

Reference Standard Establishment

Manual radiographic measurement and Lenke classification were independently performed by 2 spine surgeons according to the standard Lenke criteria. The proximal thoracic, main thoracic, and thoracolumbar/lumbar Cobb angles, T5–T12 kyphosis, and modifier-related radiographic features were assessed on standing and bending radiographs. Disagreements were resolved by review from a senior spine surgeon, and the final reference standard was determined by consensus.

Vertebral axial rotation was additionally graded on the standing posteroanterior radiographs using the NashMoe method. ²⁷ Rotation was classified as grades I–IV according to the progressive medial displacement of the convex-side pedicle relative to the vertebral body: grade I, the pedicle migrates within the lateral portion of the vertebral body; grade II, the pedicle reaches the middle third; grade III, the pedicle approaches or reaches the midline; and grade IV, the pedicle crosses beyond the midline. Rotation grades were assigned independently by the same two spine surgeons, and disagreements were resolved by consensus with the senior surgeon.

For clinical applicability analysis, 20 randomly selected cases from the independent test set were further reviewed by 3 spine surgeons with different levels of experience. Each surgeon independently assessed the cases, and agreement with the AI system as well as review efficiency under conventional and AI-assisted workflows was evaluated.

Statistical Analysis for Performance Evaluation

All statistical analyses were performed using R software. Continuous variables were summarized as mean ± standard deviation or median with interquartile range, as appropriate, and categorical variables were summarized as counts and percentages. For radiographic measurements, agreement between AI-derived and expert reference measurements was evaluated using mean absolute error (MAE), intraclass correlation coefficient (ICC), Pearson correlation coefficient, and Spearman rank correlation coefficient. For classification tasks, performance was evaluated using accuracy for Lenke curve type, lumbar modifier, sagittal thoracic modifier, and complete Lenke label assignment. Agreement between the AI system and clinicians, as well as interobserver agreement among clinicians, was assessed using Cohen’s kappa. Efficiency outcomes were summarized descriptively by comparing conventional manual review and AI-assisted review workflows. A 2-sided P value < 0.05 was considered statistically significant where applicable.

Results

Model Convergence

The proposed framework was evaluated in terms of training convergence, final performance on the independent test set, and agreement with clinicians. The training dynamics of the optimized vertebral detection and keypoint detection models are shown in Figure 2. Both models exhibited stable convergence, with only a small gap between training and validation performance, suggesting satisfactory generalization. Based on the final converged models, the framework was further evaluated on the independent test set with respect to vertebral detection, keypoint localization, Cobb angle measurement, and automated Lenke classification.

Figure 2.

Training and validation curves of the optimized vertebral detection and keypoint detection models. The green dashed lines indicate the best checkpoints

Independent Test-Set Performance

The independent test set consisted of 76 AIS cases that were not used for training or validation. Gold-standard Lenke labels were established independently by two senior spine surgeons. The test set included 28 Type 1 cases, 12 Type 2 cases, 14 Type 3 cases, 6 Type 4 cases, 10 Type 5 cases, and 6 Type 6 cases. The final performance of the proposed framework on the independent test set is summarized in Table 1.

Table 1.

Final Performance of the Proposed Framework on the Independent Test Set

Metric	Value	Prespecified criterion
Independent test cases, n	76	–
Vertebral detection AP@0.75	0.858	≥0.85
Overall keypoint AP	0.951	≥0.95
Cobb angle MAE (°)	2.4	≤3.0
Cobb angle ICC	0.991	≥0.95
AI–expert Pearson correlation	0.987	≥0.95
AI–expert Spearman correlation	0.983	≥0.95
Lenke curve-type accuracy	0.895	≥0.85
Lumbar modifier accuracy	0.868	≥0.80
Sagittal thoracic modifier accuracy	0.908	≥0.85
Complete Lenke label accuracy	0.816	≥0.75
Mean inference time per case (s)	12.3	≤30

MAE, mean absolute error; ICC, intraclass correlation coefficient.

Before consensus, independent assessments by the two surgeons demonstrated strong reliability (ICC ≥ 0.928 for Cobb and kyphosis angles; Cohen’s κ ≥ 0.768 for Lenke classifications). Minor disagreements in thresholdadjacent cases highlight the inherent subjectivity of manual classification, justifying the proposed automated framework.

On the independent test set, the system achieved a vertebral detection AP@0.75 of 0.858 and an overall keypoint AP of 0.951. The mean absolute error of Cobb angle measurement was 2.4°. Agreement between AI-derived and expert measurements was high, with an intraclass correlation coefficient of 0.991, a Pearson correlation coefficient of 0.987, and a Spearman rank correlation coefficient of 0.983. For classification, the proposed framework achieved accuracies of 0.895 for Lenke curve type, 0.868 for the lumbar modifier, and 0.908 for the sagittal thoracic modifier, yielding a complete Lenke label accuracy of 0.816. To characterize the precision of these estimates, Wilson 95% confidence intervals (CIs) were computed: 0.806–0.946 for Lenke curve type, 0.774–0.927 for the lumbar modifier, 0.822–0.955 for the sagittal thoracic modifier, and 0.714–0.887 for the complete Lenke label. The mean inference time per case was 12.3 seconds. Collectively, these results indicate that the proposed framework achieved strong performance across detection, measurement, and classification tasks.

Subgroup Analysis

A subgroup analysis was performed for cases with Cobb angles exceeding 60° (n = 11) and for cases with NashMoe Grade III/IV axial rotation (n = 18). Curve-type accuracy was 0.727 (8/11) in the high-angle subgroup and 0.722 (13/18) in the high-rotation subgroup. Given the small subgroup sizes, particularly for cases meeting both criteria (n = 3), these estimates carry wide confidence intervals and are presented descriptively rather than as formal comparative statistics.

Clinician-Referenced Workflow Evaluation

To further assess clinical applicability, three spine surgeons with different levels of experience independently reviewed 20 randomly selected test cases. The agreement between the AI system and each clinician, as well as the corresponding efficiency comparison, is shown in Table 2.

Table 2.

Agreement Between the AI System and Clinicians, With Efficiency Comparison

Comparison/metric	Lenke curve-type κ	Complete-label κ	Remarks
AI vs Surgeon A	0.873	0.781	Senior surgeon
AI vs Surgeon B	0.856	0.764	Intermediate seniority
AI vs Surgeon C	0.841	0.738	Junior surgeon
Surgeon A vs Surgeon B	0.891	0.802	Senior–senior agreement
Surgeon A vs Surgeon C	0.834	0.751	–
Surgeon B vs Surgeon C	0.812	0.723	–
Efficiency metric	Manual	AI-assisted	Improvement
Mean time per case (min)	8.2	1.5	5.5×
Total time for 20 cases (min)	164	30	5.5×
Mean Cobb angle measurement time (s/angle)	45	0	Fully automated

Surgeon A, chief physician; Surgeon B, associate chief physician; Surgeon C, attending physician.

The AI system achieved a Cohen’s κ of 0.873 for Lenke curve type and 0.781 for the complete Lenke label when compared with the most senior surgeon (Surgeon A). Notably, this level of agreement approached the interobserver agreement between the two senior surgeons (Surgeon A vs Surgeon B, κ = 0.891 for curve type and κ = 0.802 for the complete label). Agreement between the AI system and the other two surgeons also remained high. In parallel, the AI-assisted workflow substantially improved efficiency: the mean review time decreased from 8.2 minutes to 1.5 minutes per case, and the total time required for 20 cases decreased from 164 minutes to 30 minutes. In addition, Cobb angle measurement, which required an average of 45 seconds per angle under manual assessment, was fully automated in the proposed system. These findings indicate that the proposed framework provided not only high diagnostic consistency but also a marked reduction in clinical workload.

Error-Based Performance Across Lenke Types

To characterize the remaining classification errors, we further analyzed the confusion matrix of Lenke curve-type prediction on the independent test set, as shown in Figure 3. Most cases were correctly classified along the diagonal, indicating good overall discriminative performance.

Figure 3.

Confusion matrix of Lenke curve-type classification on the independent test set. Darker diagonal cells indicate correctly classified cases

Only 8 curve-type misclassifications were observed among the 76 test cases: two Type 2 cases were misclassified as Type 1, two Type 1 cases as Type 3, two Type 4 cases as Type 2, and two Type 6 cases as Type 5. Review of these cases showed that they were all borderline cases near clinically relevant decision thresholds, with representative examples including a proximal thoracic Cobb angle of 26°, a thoracolumbar Cobb angle of 24°, a post-bending thoracolumbar Cobb angle of 26°, and a main thoracic Cobb angle of 23°.

Across the entire test set, the flag was triggered in 46 of 76 cases (60.5%). Notably, all 8 misclassified cases belonged to this flagged group, whereas all 30 unflagged cases were classified correctly. The flag therefore reliably separates the subset of cases requiring manual verification from those in which the automated result can be accepted with high confidence, supporting the use of the system as an autonomous first-pass screening tool with a built-in safety net. These findings confirm that the residual errors were primarily related to threshold-adjacent clinical decisions rather than gross anatomical recognition failure.

Discussion

In the present study, we developed and validated a fully automated framework for Lenke classification in adolescent idiopathic scoliosis (AIS) based on multi-view full-spine radiographs. The proposed method was able to complete vertebral localization, landmark estimation, Cobb angle measurement, and automated determination of the Lenke curve type, lumbar modifier, and sagittal thoracic modifier within a unified workflow. On the independent test set, the framework achieved favorable performance across multiple tasks, including vertebral detection, keypoint localization, radiographic measurement, and final classification. In addition, the agreement between the AI system and senior spine surgeons was high, indicating that the proposed framework has the potential to provide clinically meaningful assistance in preoperative radiographic assessment of AIS.

A major clinical challenge in the preoperative assessment of AIS is the inherent diagnostic discrepancy between spine surgeons at top-tier clinical centers and those at secondary medical centers. Lenke classification relies heavily on a sequence of complex, experience-dependent rule-based judgments, particularly when evaluating curve structurality on bending films or threshold-adjacent modifiers. Surgeons at secondary or regional centers, who may encounter severe or complex AIS cases less frequently, might exhibit greater variability in these assessments. The proposed fully automated framework directly addresses this pain point by standardizing the measurement and classification process. By deploying this model as an accessible, cloud-based decisionsupport tool, it can empower surgeons at secondary centers to perform preoperative radiographic evaluations with the accuracy and consistency of experienced specialists. This not only mitigates diagnostic discrepancies but also facilitates more standardized surgical planning across different tiers of the healthcare system.

The clinical value of the present study lies primarily in its ability to standardize one of the most experiencedependent components of AIS evaluation, namely radiographic measurement and Lenke classification. In routine practice, Lenke classification is not determined by a single global descriptor, but rather by a sequence of interrelated measurements and rule-based judgments, including curve magnitude, coronal balance, sagittal alignment, and structurality on bending views. As a result, small deviations in any intermediate step may propagate to the final classification outcome. By integrating vertebral localization, landmark estimation, radiographic parameter computation, and rulebased classification into a single automated pipeline, the present framework reduces the need for repeated manual measurement and may help improve consistency in the overall classification process.^23,28,29

Another important feature of this study is the modular design of the proposed framework. Rather than directly generating a classification label from the whole radiograph in an end-to-end black-box manner, the framework follows a clinically interpretable sequence of vertebral detection, landmark estimation, radiographic parameter computation, and final Lenke assignment. This design preserves explicit intermediate anatomical and geometric information, including vertebral centers, vertebral corners, pedicle-related landmarks, the S1 midpoint, Cobb angles of different curve segments, CSVL-related coronal alignment, and T5–T12 kyphosis. Such explicit intermediate outputs improve the interpretability of the model and make the automated decision pathway more consistent with the actual clinical workflow used in manual Lenke classification.^18,24

Beyond classification accuracy, a central question for any Lenke classification tool is how its output translates into operative decision-making, because the Lenke system was originally designed to determine the extent of spinal arthrodesis. ⁷ In this regard, the present study was deliberately scoped to the radiographic measurement and classification task: we validated that the automated pipeline reproduces expert Lenke labels and measurements with high fidelity, which is the prerequisite for any downstream surgical-planning use. Once a Lenke label is established, it maps onto well-described fusion-selection templates (e.g., selective thoracic fusion for Lenke 1/2 curves with a nonstructural lumbar modifier, and fusion of all structural curves for double/triple-major patterns), ⁶ providing an explicit bridge from the automated classification to a preliminary fusion-level proposal.

The observed reduction in review time (8.2 to 1.5 min per case) may therefore be interpreted not merely as a workflow convenience but as a means of freeing the surgical team to focus on the operative decision itself, including the extent of instrumentation, the number and distribution of screws, and the need for additional releases such as Ponte osteotomies. Whether the automated label can serve as a reliable basis for these specific operative parameters—screw density, implant distribution, osteotomy decisions, and correction maneuvers—cannot be answered by the present radiographic-only dataset and requires prospective validation against actual surgical plans and outcomes, which we have defined as the next phase of this work.

The results also suggest that the main source of residual error in the current system is not gross anatomical recognition failure, but rather borderline cases near clinically relevant decision thresholds. In the independent test set, the remaining curve-type misclassifications occurred in cases with Cobb angle measurements close to the thresholds used for structurality determination or modifier assignment. This finding is clinically meaningful because it indicates that the difficulty of automated Lenke classification does not solely arise from image understanding, but also from the high sensitivity of rule-based categorization to small geometric differences.^28,29 For this reason, the current framework may be better positioned as a decision-support tool rather than a fully autonomous replacement for clinical judgment.

In particular, the automatic review flag for threshold-adjacent cases may be useful in practice, as it allows the system to provide rapid and consistent results in standard cases while prompting additional human verification in more ambiguous situations.

The present study further demonstrated that model optimization improved performance not only at the overall level, but also in relatively uncommon Lenke subtypes. This observation is important because the practical challenge of automated Lenke classification lies not only in achieving a high overall accuracy, but also in maintaining stable performance under class imbalance. In our results, the final model showed improved recall for less frequent subtypes, suggesting that the framework retained reasonable robustness even in categories that are typically more difficult to classify. From a clinical perspective, this is relevant because rare or complex curve patterns are often the cases in which consistent preoperative classification is most needed.

Several limitations of the present study should be acknowledged. First, despite the use of an independent multicenter test set and expert consensus annotations, the overall sample size remains modest (n = 76), and the distribution of Lenke subtypes is inherently imbalanced. For the least frequent curve patterns (e.g., Types 4 and 6, n = 6 each), the Wilson 95% confidence intervals for sensitivity were wide (0.300–0.903), reflecting limited statistical power rather than necessarily inferior model performance; expanding the test cohort for these underrepresented subtypes is an ongoing priority.

Conclusion

A fully automated deep learning framework for Lenke classification in adolescent idiopathic scoliosis was developed and validated using multi-view full-spine radiographs. The method achieved reliable automated radiographic measurement and Lenke classification, while maintaining strong agreement with experienced clinicians. These findings support its potential as an interpretable and efficient decisionsupport tool for preoperative AIS assessment.

ORCID iDs

Sheyang Xu https://orcid.org/0009-0002-1809-4271

Yongda Xu https://orcid.org/0009-0001-9083-0045

Ethical Considerations

As this study utilized retrospective data analysis, a waiver of informed consent was granted by the ethics committee in accordance with national regulations and the Declaration of Helsinki.This study was approved by the Ethics Committee of Beijing Anzhen Hospital (Ethical Approval No. 2026122x).

Data Availability Statement

Data availability is not applicable to this study.*