Can classification strategies improve automated cervical vertebral maturation staging? A comparative study.

Accurate assessment of skeletal maturity is essential in determining orthodontic treatment timing. Cervical vertebral maturation (CVM) staging is commonly used, but inter-observer variability remains a major obstacle. Recently, artificial intelligence (AI) has been used to provide more consistent and faster radiographic assessments. This study aimed to compare various deep learning approaches and investigate how different training strategies impact model performance in automated CVM staging. A total of 1,750 lateral cephalometric radiographs were independently evaluated and labeled by two board-certified orthodontists, with discrepancies resolved by consensus. This dataset was then divided into a training set (n = 1,600; stratified 5-fold cross-validation) and a held-out test set (n = 150). First, we compared the end-to-end 6-stage model (LS6) with landmark-guided 6-stage models (LM6_1 and LM6_2, collectively referred to as LM6) to test the effect of structural priors. Then, we evaluated a fine-to-coarse 3-stage model (LS6_3), trained on 6 stages and then aggregated for 3-stage classification, against a direct 3-stage model (LS3) to assess the impact of label granularity. For 6-stage classification, LS6 achieved 67.3% accuracy (κ = 0.912), outperforming LM6_1 (58.8%) and LM6_2 (64.4%). Tolerance-based analysis demonstrated high ± 1-stage accuracy for all models, with LS6 achieving 94.3% and LM6 models achieving 92.9% (LM6_1) and 92.8% (LM6_2), confirming that misclassifications were predominantly between adjacent stages. For 3-stage classification, LS6_3 (79.3%) showed slightly higher accuracy than LS3 (78.8%). In the Grad-CAM analysis, LS6_3 showed a more concentrated focus on key vertebral features compared to LS3. These findings suggest that landmark-based priors may not necessarily enhance CVM staging performance, and that fine-grained training encourages more anatomically focused feature learning. These results provide preliminary evidence for optimizing AI training protocols in skeletal maturity assessment.