<- Back to digest

JBJS - 2026-06-17 - Journal Article

A Fully Automated Multistage Deep Learning System for Lenke Classification: Enhanced Diagnostic Precision in Adolescent Idiopathic Scoliosis.

Xu L, Liu C, Zhong W, Xu K, He N, Hu J, Zhang K, Cui H

retrospective cohortLOE IIIn = 650 (467 AIS, 183 controls)N/A

Topics

pediatricsspine
PMID: 41460944DOI: 10.2106/JBJS.25.01015View on PubMed ->

Key Takeaway

A fully automated deep learning system achieved 95.6% overall accuracy and macro-averaged F1 score of 0.862 for Lenke classification of AIS directly from spinal radiographs, with Cobb angle ICC of 0.969–0.976 versus expert measurement.

Summary Depth

Choose how much analysis to show on this article page.

Summary

This study asked whether a fully automated multistage deep learning pipeline could perform end-to-end Lenke classification from full-spine radiographs with expert-level accuracy. The system used Swin-Unet vertebral segmentation, DeepLabv3+ lumbar pedicle localization, and a fusion classification module trained and validated on 650 subjects aged 10–18 with Cobb angles 25–89°. Overall Lenke classification accuracy was 95.6%, lumbar modifier F1 was 0.912, sagittal modifier F1 was 0.928, and Cobb angle agreement with experts yielded ICC 0.969–0.976.

Key Limitation

The dataset is 95% Han Chinese from a single center, and performance on Western or ethnically diverse AIS populations with different curve morphology distributions has not been established.

Original Abstract

BACKGROUND

The Lenke classification for adolescent idiopathic scoliosis (AIS) has interobserver variability due to subjective clinical assessment. We developed and validated a fully automated deep learning system for precise Lenke classification using spinal radiographs.

METHODS

This retrospective study included 650 individuals (mean age, 13.75 ± 2.23 years; 433 female, 217 male; 618 Han Chinese, 32 Tibetan), comprising 183 healthy controls and 467 patients with AIS (aged 10 to 18 years; 25° ≤ Cobb angle < 90°) with full-spine radiographs. A multistage deep learning system consisting of (1) Swin-Unet segmentation of vertebrae (C7-S1) for automated Cobb angle measurement, (2) DeepLabv3+ localization of lumbar pedicles (L1-L5) to determine modifiers via the centroid-to-CSVL (central sacral vertebral line) distance, and (3) a fusion module integrating features to curve types and lumbar (A/B/C) and sagittal thoracic (-/N/+) modifiers was designed to perform end-to-end Lenke classification automatically. Validation used an independent test set.

RESULTS

The system achieved 95.6% overall accuracy in Lenke classification and had a macro-averaged F1 score of 0.862. Vertebral segmentation attained Dice coefficients of 0.917 (anteroposterior) and 0.942 (lateral). Cobb angle measurements showed excellent agreement with those of experts (intraclass correlation coefficient, 0.969 to 0.976 for thoracic or thoracolumbar/lumbar curves). Modifier assignment achieved F1 scores of 0.912 (lumbar A/B/C) and 0.928 (sagittal -/N/+), exceeding clinical acceptability thresholds.

CONCLUSIONS

The fully automated system was able to perform rapid, objective, interpretable, and clinically reliable classification of the Lenke type directly from radiographs, with performance comparable with that of expert assessment. It demonstrates potential for standardizing AIS surgical planning, reducing diagnostic variability, and improving surgical workflow efficiency.

LEVEL OF EVIDENCE

Diagnostic Level III . See Instructions for Authors for a complete description of levels of evidence.