Blog
AggregateNet: A deep learning model for automated classification of cervical vertebrae maturation stages
This study presents AggregateNet, a novel deep learning architecture designed to fully automate the classification of cervical vertebrae maturation (CVM) stages using lateral cephalometric radiographs. CVM staging is a crucial diagnostic indicator for assessing skeletal maturity in orthodontic treatment planning, but manual evaluation remains subjective, operator-dependent, and requires significant expertise. The authors address these limitations by introducing a parallel-structured CNN enhanced with directional edge filters and chronological age input. A total of 1018 cephalometric radiographs from the AAOF Craniofacial Growth Legacy Collection were included, each labeled into six CVM stages (CS1–CS6) by an expert orthodontist with high intra- and inter-examiner reliability. Images were automatically cropped to isolate the C2–C4 vertebrae, resized to standardized dimensions, and separated by gender to improve physiological modeling. Data augmentation methods (translation, rotation, autocontrast) were applied to reduce overfitting. AggregateNet incorporates three parallel sub-networks within each aggregation block, inspired by ResNeXt’s aggregated residual connections but without explicit skip connections. A preprocessing layer applies directional high-pass edge-enhancing filters to highlight subtle morphological changes in vertebral borders—critical for distinguishing adjacent CVM stages. Chronological age is concatenated with the learned feature vector in the classification head, further enhancing prediction accuracy. The model outperformed multiple baseline architectures (ResNet20, MobileNetV2, Xception, CNNDF). With directional filters + age input + augmentation, AggregateNet reached a validation accuracy of 82.35% (females) and 75.0% (males). Removing directional filters or age input reduced performance, confirming their contribution. Confusion matrices reveal that misclassifications occur primarily between adjacent stages, consistent with clinical expectations. Overall, AggregateNet demonstrates strong potential as a clinically useful tool for automated skeletal maturity assessment, reducing reliance on radiologist expertise while improving reproducibility. Its design emphasizes anatomical interpretability and multi-input synergy, representing a meaningful step toward AI-assisted orthodontic diagnostics.
artificial intelligence; cervical vertebrae maturation; CVM; growth and development; deep learning
Year: 2023
Modality: Lateral cephalometric radiographs
Dataset: 1018 images
Dataset Split: Gender-specific — 430 train / 104 test (male); 393 train / 85 test (female)
Architecture: AggregateNet (parallel sub-networks + directional edge filters + age input)
Metrics: Accuracy, confusion matrix
AP Qty: 2
Supervisor: No information
Experience: >10 years
Expertise: Orthodontist and GD
Tool/System: No information
ML Task: Multi-class classification
Objective: Automate CVM stage determination using a deep learning model with enhanced edge detection and multi-input fusion.
A reliable automated solution for CVM staging improves diagnostic consistency and reduces dependency on specialist expertise, supporting AI-assisted orthodontic planning.