Quantification of Craniofacial Growth Pattern Based on Deep Learning

Ziyi Hu, Yuyanran Zhang, Ningtao Liu, Xin Gao, Wei Fang, Guanglin Wu, Zhiyong Zhang, S W WangBioengineering 2026 · DOI

TL;DR

This paper introduces an end-to-end deep learning framework that automatically quantifies craniofacial growth patterns and sexual dimorphism from lateral cephalometric radiographs. By avoiding manual annotations and leveraging techniques like Grad-CAM, the model generates novel "saliency maps" and indices (ASI, SSI) that offer objective, detailed insights into how facial structures develop and differ between sexes over time.

What Problem Does This Paper Address?

Understanding craniofacial growth is crucial for orthodontics and maxillofacial surgery. Traditional methods rely on manual landmarking on X-rays, which is time-consuming, subjective, and simplifies complex 3D morphology into 2D points. While deep learning has shown promise in medical imaging, many approaches still require manual annotations for training, perpetuating subjectivity and limiting scalability. This study aims to overcome these limitations by developing an automated, objective system to quantify craniofacial development and sex-specific differences using a large dataset of X-rays.

Methods

The core of this study is an end-to-end deep learning framework built upon the EfficientNet-B0 architecture. This model is trained for two proxy tasks: age estimation and sex classification. By learning to predict age and sex from raw lateral cephalometric radiographs (LCRs), the network implicitly learns to identify features related to craniofacial growth and sexual dimorphism.

To interpret what the model has learned, the researchers employed Gradient-weighted Class Activation Mapping (Grad-CAM). This technique generates "saliency maps" that highlight the image regions most influential for the model's predictions. These maps are then aggregated to create population-averaged saliency maps for different age groups and sexes.

To quantify these findings, two novel metrics were introduced:

  • Age-related Saliency Index (ASI): Measures the importance of specific anatomical regions for age estimation, indicating their contribution to developmental changes.
  • Sex-related Saliency Index (SSI): Measures the importance of specific anatomical regions for sex classification, highlighting areas with significant sexual dimorphism.

Datasets

The study utilized a large dataset of 41,625 lateral cephalometric radiographs (LCRs) from individuals aged 4 to 18 years obtained from the Xi'an Jiaotong University Stomatological Hospital. The dataset was anonymized, and only age and sex labels were retained. The data was divided into training, validation, and testing sets with a 7:1.5:1.5 ratio. For sex classification analysis, specific age groups (6, 9, 12, 15, and 18 years) were selected to examine developmental stages.

Key Insights and Ideas

  • Automated Feature Extraction: The deep learning approach bypasses manual landmarking, allowing the model to discover complex, multi-tissue features relevant to growth and dimorphism that might be missed by traditional methods.
  • Visualizing Developmental Trajectories: Grad-CAM saliency maps provide an intuitive, visual representation of which craniofacial regions are most dynamic at different ages, moving beyond simple contour analysis to reveal internal bone structures.
  • Quantifying Regional Importance: The ASI and SSI metrics offer a standardized, objective way to rank the developmental significance and sex-specific differences of various anatomical regions, providing quantitative data for clinical assessment.
  • Shifting Saliency Over Time: The study reveals a clear shift in developmental saliency from the cranial region in early childhood to more localized areas like the maxilla and mandible during adolescence, offering insights into the temporal dynamics of growth.
  • Evolution of Sexual Dimorphism: The SSI analysis demonstrates how sex differences, initially widespread, become more concentrated in specific regions (like the mandible) as individuals mature.

Key Results

  • Age Estimation Performance: The EfficientNet-B0 model achieved a mean absolute error (MAE) of 0.6447 years and a coefficient of determination (R²) of 0.9444, indicating high accuracy and stability in predicting age from LCRs.
  • Sex Classification Performance: The model demonstrated exceptional performance across all age groups, with accuracy, precision, recall, and F1 scores consistently exceeding 99.5%.
  • Saliency Map Findings:
    • Early development (4-7 years) showed high saliency in the sphenoid, temporal bones, and dentition.
  • After age 8, saliency increased rapidly in the maxilla and zygoma, while the mandible also showed high saliency.
  • By age 18, saliency was high across most craniofacial areas, with males showing increased saliency in the mandibular and zygomatic regions.
  • Sexual Dimorphism Patterns:
    • At age 6, differences were concentrated in the mandible and maxilla.
  • By age 18, the mandible showed the most significant sex-related difference, with males exhibiting higher SSI.

Limitations and Future Work

  • Retrospective, Single-Center Dataset: The model's generalizability may be limited due to the retrospective nature and single-center origin of the dataset. Future work should involve multi-center, multi-ethnic cohorts.
  • 2D Radiographs: Reliance on 2D LCRs inherently limits the assessment of transverse dimensions and can lead to projective superimposition. Future research should integrate three-dimensional imaging (e.g., CBCT) to provide more comprehensive spatial information.
  • Interpreting Saliency: While Grad-CAM highlights influential regions, the precise anatomical or radiological correlates of these high-saliency features (e.g., bone density, trabecular architecture) require deeper investigation.
  • Future Directions: The authors suggest validating the model with larger, diverse datasets, exploring 3D imaging, and correlating deep learning saliency features with radiomic or biomechanical features to enhance interpretability and clinical application.

Summary updated

Abstract

Background: Childhood and adolescence constitute a critical period for craniofacial growth. Understanding its developmental patterns is essential for clinical decision-making in orthodontics and maxillofacial surgery. Traditional cephalometric analysis relies on manual landmarking, which oversimplifies complex morphology and introduces subjectivity. Although deep learning, a key artificial intelligence (AI) technology, has demonstrated remarkable performance in image analysis and classification, most methods still depend on manual annotations during training, perpetuating subjectivity and limiting model generalizability and robustness on large datasets. This hinders the development of objective, comprehensive methods to quantify craniofacial growth that account for its multi-tissue complexity. Methods: To address these limitations, this study developed an end-to-end deep learning framework based on lateral cephalometric radiographs from 41,625 individuals aged 4–18 years. Without relying on manual annotations, the model is designed to autonomously extract dynamic imaging features associated with continuous age intervals in craniofacial development and further discern features related to sexual dimorphism. Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize the learned features, generating population-averaged saliency maps that highlight age-related and sex-related patterns. Furthermore, we introduced two novel quantitative metrics, the Age-related Saliency Index (ASI) and the Sex-related Saliency Index (SSI), to evaluate the significance of developmental and dimorphic characteristics in key craniofacial regions. Results: Age-related saliency maps extended the focus from external contours to internal anatomical details of the bones, intuitively visualizing the relative importance of multiple bone regions during dynamic development, with the ASI providing a quantitative prioritization of these regions. The Sex-related Saliency Index (SSI) quantified the dynamic evolution of sexual dimorphism, demonstrating that early-stage differences were widely distributed across cranial bones and gradually became concentrated in the mandibular region by adulthood. Conclusions: This study established an end-to-end deep learning framework for analyzing large-scale lateral cephalometric radiographs. By generating age- and sex-related average saliency maps and their corresponding quantitative indices, we visualized and quantified the spatiotemporal growth dynamics and sexual dimorphism across distinct craniofacial skeletal regions throughout development. These findings not only validate established developmental theories but also provide novel insights into the coordinated growth patterns of craniofacial bones and sex-specific radiological characteristics, offering clinicians objective quantitative references for assessing developmental stages and guiding the timing of interventions targeting specific craniofacial regions.
BibTeX
@inproceedings{hu2026quantification,
  author = {Ziyi Hu and Yuyanran Zhang and Ningtao Liu and Xin Gao and Wei Fang and Guanglin Wu and Zhiyong Zhang and S W Wang},
  title = {{Quantification of Craniofacial Growth Pattern Based on Deep Learning}},
  booktitle = {Bioengineering},
  year = {2026},
  doi = {10.3390/bioengineering13030277},
}
Figure