CDSNet: an Automated Method for Assessing Growth Stages from Various Anatomical Regions in Lateral Cephalograms Based on Deep Learning
Yuchen Zhang,Zhen Lu,Jianglin Zhou,Yi Sun,Wuci Yi,Juan Wang,Tianjing Du,Dongning Li,Xinyan Zhao,Yifei Xu,Chen Li,Kun Qi
DOI: https://doi.org/10.1016/j.ejwf.2024.09.007
2024-01-01
Journal of the World Federation of Orthodontists
Abstract:Background The assessment of growth stages, typically determined by Cervical Vertebrae Maturation (CVM), plays a crucial role in orthodontics. However, there is a potential deviation from actual growth stages when using CVM. This study aimed to introduce CDSNet, an interpretable deep learning model for assessing growth stages based on cervical vertebrae, dentition, and frontal sinus in lateral cephalograms. Methods A dataset of 1,732 pairs of lateral cephalograms and hand-wrist radiographs from patients who underwent orthodontic treatment was annotated by four dentists. Benchmarks were conducted using CVM and logistic regression. Experiments were designed to evaluate CDSNet's performance in assessing growth stages using various methods and anatomical regions. Results CDSNet achieved remarkable Accuracy (90.99%), Precision (89.98%), Recall (92.50%), and F-1 Score (91.22%) in assessing growth spurt, representing significant improvements of 26.56%, 27.96%, 30.26%, and 29.30% compared to the CVM-based method. Additionally, when compared to a deep learning method based on cervical vertebrae, improvements of 12.25%, 11.40%, 14.14%, and 12.56% were observed. The interpretable module's side output revealed the involvement of cervical vertebrae, dentition, and frontal sinus in assessing growth spurt. Conclusions In the clinical domain, CDSNet is able to assist clinicians in determining patients’ growth stages, particularly those near the boundary between two stages with less distinct features. This study demonstrated the role of interpretable deep learning in investigating the external manifestations of craniofacial growth. Integrating algorithmic or clinical research to analyze multiple features on lateral cephalograms proved a feasible approach to assist orthodontists and improve diagnostic efficacy.