Service Support

Recently, the team of Prof. Zhao Changqing, Department of Orthopedics, Ninth People's Hospital, Shanghai Jiao Tong University School of Medicinepublished a research paper in the Journal of Spinal Surgeryentitled "Validation Study of an AI Automatic Lenke Classification Model for Adolescent Idiopathic Scoliosis". This study used SiPuNai Medical Technology (Shanghai) Co., Ltd.developed the Ruiji (SpineSight) image processing software to build an AI model for measuring AIS patients' Cobb angles and performing Lenke classification, evaluating its feasibility and practicality in clinical practice.

 

This study developed an AI model to automatically measure the Cobb angle of adolescent idiopathic scoliosis (AIS) and perform Lenke classification, aiming to improve measurement efficiency and consistency. Although AI methods (e.g., CNN) achieve acceptable consistency with manual measurements, clinical adoption still faces challenges such as "black-box diagnosis" and medical liability. The study emphasizes that enhancing algorithm interpretability and allowing manual verification and adjustment can increase clinicians' trust in automated measurement and promote its application in practice.

1. Literature Search

Inclusion criteria: 1) clear AIS diagnosis with definite Lenke classification; 2) patient age 10–20 years; 3) clear and complete imaging data.

Exclusion criteria: 1) poor image acquisition or stitching quality; 2) severe pelvic tilt; 3) previous spinal surgery history.

Based on the above criteria, 7 publications with 14 cases were finally selected (1 male, 13 female; age 14.88±2.55 years). All imaging data were used as the validation set to evaluate the accuracy of Cobb angle measurement and Lenke classification.

2. Model Construction

2.1. Landmark Detection

Based on the Ruiji (SpineSight) image processing software, this study developed an AI model for automatically measuring spinal parameters. The model uses an asymmetric encoder-decoder network architecture. The encoder is based on the RepVGG-A1 architecture, which introduces ResNet-like shortcut connections during training to ensure smooth gradient propagation, while being equivalent to an efficient VGG-style convolutional network at inference. With approximately 13M parameters, the model balances powerful feature extraction with real-time inference speed.

The model performs two key tasks: generating pixel-wise segmentation masks of the vertebrae and predicting the offset between each vertebral pixel and its four corner landmarks. During training, the segmentation task is supervised with binary cross-entropy loss, while the corner regression task is optimized with smooth L1 loss. At inference, the algorithm combines the segmentation mask and predicted offsets to compute precise vertebral corner coordinates.

Performance validation showed that the model's mean radial error (MRE) in locating vertebral landmarks on standing full-length spine X-rays was only 1.24 mm on anteroposterior films and 1.28 mm on lateral films, with successful detection rates (SDR) within 2 mm reaching as high as 90.46% and 88.93%, respectively.In addition, the Dice similarity coefficient (DSC) reached 0.83 for pedicle segmentation, further confirming the model's high accuracy and clinical practicality in spinal image analysis.

2.2. Cobb Angle Measurement

This study Lenke classificationdetermined the Cobb angle measurement regions based on Lenke classification, identified vertebral landmarks on full-length anteroposterior and lateral bending X-rays with the AI model, and automatically calculated the Cobb angle of each region. Bland-Altman analysis was used to evaluate agreement. Results showed excellent agreement between automated and manual measurements (ICC>0.86), with small mean absolute error and root mean square error for the major curve Cobb angle, and measurement differences within the clinically acceptable range of 5°, demonstrating the reliability and practicality of the automated method in classification and measurement.

2.3.Lenke classification

This study used the AI model to automatically measure Cobb angles and sagittal parameters, performing Lenke main classification and modifier classifications of lumbar coronal (A/B/C) and thoracic sagittal (-/N/+) types. Validation on external datasets showed the model achieved accuracies of 89%, 89%, and 96% on the three classifications respectively, confirming the high accuracy of the automated classification system and providing a reliable tool for clinical use.

3. Statistical Analysis

Statistical analysis was performed using SPSS, GraphPad Prism, and PyCharm software. After Shapiro-Wilk normality testing, measurement data were expressed as mean ± standard deviation, and between-group comparisons were made with t-tests. Reliability was assessed using the intraclass correlation coefficient, absolute agreement was measured by mean absolute difference and its standard deviation, and Bland-Altman plots were used to analyze agreement and systematic bias. The classification performance of the AI model was evaluated using confusion matrix metrics (accuracy, precision, recall, etc.). Differences were considered statistically significant at P<0.05.