Scholarlink

Introduction <p>Traditional orthodontic diagnostics rely significantly on lateral cephalometric radiographs, posing health risks due to ionising radiation, particularly in paediatric patients. Artificial intelligence (AI) represents a promising alternative by enabling predictions of cephalometric parameters from non-radiographic clinical data. This study evaluates the accuracy and clinical utility of CEPHCLINIC, an open-source AI software designed to predict conventional cephalometric measurements using clinical photographs and intraoral 3D scans, thus adhering to radiation protection principles.</p> Materials and methods <p>The dataset comprised 1255 subjects from the American Association of Orthodontists Foundation (AAOF) craniofacial collection, encompassing demographic and clinical variables (age, gender, overbite, overjet, facial dimensions). This dataset was randomly divided into training (80%, <i>n</i> = 1004) and validation (20%, <i>n</i> = 251) subsets. Additionally, an independent external test set of 51 untreated orthodontic cases was employed for rigorous evaluation. Input variables for model training included clinical parameters derived from photographs (WebCeph software) and intraoral scans (iTero scanner, MeshMixer software). Supervised predictive regression models, including ExtraTreesRegressor, CatBoostRegressor, and Support Vector Regression, were optimised through GridSearchCV and validated using repeated random subsampling. Predictive accuracy was assessed statistically using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Spearman correlation coefficients, R-squared values, and paired t-tests.</p> Results <p>The ExtraTreesRegressor demonstrated superior performance across multiple cephalometric parameters, achieving notably low RMSE values in the independent test set for ANB (2.772°) and NP2PA (2.317&#xa0;mm). However, parameters like COPAD exhibited higher prediction errors (RMSE 12.121&#xa0;mm). Spearman correlation analysis indicated strong prediction consistency for COPOD (0.850), moderate for U1SNA (0.548), and poor predictability for NP2PO (-0.052). Despite statistically significant biases observed in predictions for some parameters (paired t-test, <i>p</i> &lt; 0.05), overall predictive accuracy was clinically acceptable, emphasising parameters such as COPOD, COPAD, ANB, and U1SNA as particularly reliable.</p> Conclusions <p>The CEPHCLINIC software reliably predicts key cephalometric measurements from non-radiographic clinical data, significantly reducing radiation exposure risks. Despite promising performance, model refinement, dataset expansion with broader demographic representation, and integration with digital orthodontic technologies are essential for enhancing precision, clinical reliability, and global applicability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An open-source AI tool for predicting cephalometric measurements from clinical data and photographic images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月3日

查看全文
  • Piero Antonio Zecca,
  • Margherita Caccia,
  • Luca Levrini,
  • Andrea Carganico,
  • Marco Serafin,
  • Petra Rita Basso,
  • Marcella Reguzzoni

摘要

Introduction

Traditional orthodontic diagnostics rely significantly on lateral cephalometric radiographs, posing health risks due to ionising radiation, particularly in paediatric patients. Artificial intelligence (AI) represents a promising alternative by enabling predictions of cephalometric parameters from non-radiographic clinical data. This study evaluates the accuracy and clinical utility of CEPHCLINIC, an open-source AI software designed to predict conventional cephalometric measurements using clinical photographs and intraoral 3D scans, thus adhering to radiation protection principles.

Materials and methods

The dataset comprised 1255 subjects from the American Association of Orthodontists Foundation (AAOF) craniofacial collection, encompassing demographic and clinical variables (age, gender, overbite, overjet, facial dimensions). This dataset was randomly divided into training (80%, n = 1004) and validation (20%, n = 251) subsets. Additionally, an independent external test set of 51 untreated orthodontic cases was employed for rigorous evaluation. Input variables for model training included clinical parameters derived from photographs (WebCeph software) and intraoral scans (iTero scanner, MeshMixer software). Supervised predictive regression models, including ExtraTreesRegressor, CatBoostRegressor, and Support Vector Regression, were optimised through GridSearchCV and validated using repeated random subsampling. Predictive accuracy was assessed statistically using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Spearman correlation coefficients, R-squared values, and paired t-tests.

Results

The ExtraTreesRegressor demonstrated superior performance across multiple cephalometric parameters, achieving notably low RMSE values in the independent test set for ANB (2.772°) and NP2PA (2.317 mm). However, parameters like COPAD exhibited higher prediction errors (RMSE 12.121 mm). Spearman correlation analysis indicated strong prediction consistency for COPOD (0.850), moderate for U1SNA (0.548), and poor predictability for NP2PO (-0.052). Despite statistically significant biases observed in predictions for some parameters (paired t-test, p < 0.05), overall predictive accuracy was clinically acceptable, emphasising parameters such as COPOD, COPAD, ANB, and U1SNA as particularly reliable.

Conclusions

The CEPHCLINIC software reliably predicts key cephalometric measurements from non-radiographic clinical data, significantly reducing radiation exposure risks. Despite promising performance, model refinement, dataset expansion with broader demographic representation, and integration with digital orthodontic technologies are essential for enhancing precision, clinical reliability, and global applicability.