Related Experiment Videos
From routine full-spine radiographs to decision-oriented Risser stratification: an interpretable deep-learning
Zexi Wang1, Yuan Zhang2, Yixi Wang1
1Department of Minimally Invasive Spine and Precision Orthopedics, The First Affiliated Hospital of Xinjiang Medical University, Urumqi, China.
Frontiers in Pediatrics
|July 2, 2026
Summary
A deep-learning model accurately stratifies skeletal maturity in adolescent idiopathic scoliosis (AIS) using radiographs, improving efficiency and aiding treatment planning. This AI tool enhances Risser staging for better growth potential assessment.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Medicine
- Orthopedics and Spine Surgery
Background:
- Accurate assessment of remaining growth is critical for risk stratification and treatment planning in adolescent idiopathic scoliosis (AIS).
- Risser staging, a common method, exhibits moderate reproducibility in clinical practice, especially on standard radiographs where key features are small.
- This limitation highlights the need for more objective and reproducible methods for skeletal maturity assessment in AIS.
Purpose of the Study:
- To develop and evaluate an interpretable deep-learning model for automatic skeletal maturity stratification in AIS.
- The model aims to classify patients into Risser stages 0-2 versus 3-5 using routine full-spine radiographs.
- To assess the model's interpretability and its impact on clinical readers' efficiency and agreement.
Main Methods:
- A retrospective study utilized 875 standing posteroanterior full-spine radiographs from AIS patients (aged 10-18 years).
- An expert consensus standard was established, and a deep learning model (ResNet-18) was trained on automatically extracted pelvic regions.
- Model performance was evaluated using AUC and Cohen's kappa, with interpretability assessed via Grad-CAM; a reader study compared unaided vs. model-assisted readings.
Main Results:
- The deep-learning model achieved high performance for binary stratification (Risser 0-2 vs. 3-5) with an AUC of 0.938 and accuracy of 0.875.
- Gradient-weighted class activation mapping (Grad-CAM) confirmed the model focused on relevant iliac apophysis ossification regions.
- Model assistance significantly reduced reading time (by 9.7-11.8 seconds per case) and improved reader agreement with expert consensus, particularly for junior surgeons.
Conclusions:
- An interpretable deep-learning model can effectively perform Risser stratification using routine radiographs, aiding clinical decision-making in AIS.
- The AI tool demonstrated improved reading efficiency and accuracy, supporting standardized assessment of growth potential without additional imaging.
- This approach holds promise for enhancing routine management of adolescent idiopathic scoliosis.
