From routine full-spine radiographs to decision-oriented Risser stratification: an interpretable deep-learning approach.

Background: Accurate assessment of remaining growth is crucial for risk stratification and treatment planning in adolescent idiopathic scoliosis (AIS). Risser staging is widely used but demonstrates only moderate reproducibility in routine practice, particularly on standing full-spine radiographs where the iliac apophysis occupies a small area. This study aimed to develop and evaluate an interpretable, decision-oriented deep-learning model for automatic stratification of skeletal maturity as Risser 0-2 vs. 3-5 using routine full-spine radiographs. Methods: This single-center retrospective study included 875 standing posteroanterior full-spine radiographs from AIS patients aged 10-18 years. An expert consensus reference standard was established using a three-surgeon adjudication process. A predefined, rule-based pelvic region of interest was automatically extracted for model training. The data were split at the patient level into training, validation, and test sets (70%/10%/20%, stratified). Multiple convolutional neural network architectures were evaluated, and ResNet-18 was selected based on performance stability. Model interpretability was assessed using gradient-weighted class activation mapping (Grad-CAM). An independent clinical reader study was conducted on 50 sampled test cases, in which three spine surgeons completed two reading sessions (without and with model assistance) to assess agreement with the expert reference and reading time. Results: For binary stratification (Risser 0-2 vs. 3-5), the selected model achieved an area under the receiver operating characteristic curve (AUC) of 0.938, an accuracy of 0.875, and a Cohen's kappa of 0.748 on the independent test set. Grad-CAM visualizations indicated that model predictions focused on the iliac apophysis and adjacent ossification regions. In the reader study, compared with unaided reading, model assistance reduced the mean reading time by 9.7-11.8 s per case across the three readers, from 34.7-43.2 s to 25.0-31.8 s (all Conclusions: Using routine standing full-spine radiographs, an interpretable deep-learning model enabled clinically actionable Risser stratification and improved reading efficiency in an independent reader study, without the need for additional imaging. This approach may support the standardized assessment of growth potential in routine AIS management.