Scholarlink

Objective <p>This study aimed to develop and validate a deep learning-based artificial intelligence (AI) system capable of automatically generating accurate MSE placement plans to enhance clinical efficiency, safety, and consistency.</p> Methods <p>A total of 120 patients with skeletal Class III malocclusion and transverse maxillary deficiency (ages 10–39) were included. Cone-beam computed tomography (CBCT) and intraoral scan data were used to construct individualized three-dimensional anatomical coordinate systems. A deep learning model was trained to automatically segment key anatomical structures (e.g., the incisive foramen and transverse palatine suture) and identify essential craniofacial landmarks. The MSE was then automatically positioned within this coordinate system, while a collision detection algorithm ensured appropriate spacing from the palatal mucosa and avoided vital anatomical structures. Model performance was compared with manual placement using metrics such as mean Intersection over Union (mIoU), Avoidance Success Rate (ASR), Mean Radial Error (MRE), Mean Angular Error (MAE), and Concordance Correlation Coefficient (CCC).</p> Results <p>The average mIoU of the segmentation network was 0.75. The ASR exceeded 90%, demonstrating effective anatomical avoidance. The automated system achieved an axial MRE of 0.32 ± 0.32&#xa0;mm, a three-dimensional (3D) Euclidean distance error of 0.69 ± 0.36&#xa0;mm, and an MAE of 1.84 ± 2.13°, with CCC values above 0.90 for linear and translational directions and between 0.87 and 0.93 for angular directions. These results showed no statistically significant differences from manual placement (<i>P</i> &gt; 0.05).Bland–Altman analysis further showed minimal mean bias between AI and manual planning (translation bias ≤ 0.047&#xa0;mm; rotation bias within − 0.26° to 1.52°), with 95% limits of agreement for all parameters remaining within clinically acceptable ranges; a mild proportional bias was observed only for yaw (<i>P</i> = 0.013).The automated process required an average of 3&#xa0;min, significantly faster than the 45–60&#xa0;minutes typically required for manual planning.</p> Conclusion <p>The proposed deep learning–based automated MSE positioning system enables accurate and efficient MSE digital planning in a retrospective setting, shows promising clinical potential, and may contribute to more standardized and intelligent orthodontic expansion workflows.</p> Clinical relevance <p>This deep learning–driven tool may facilitate standardized MSE placement, shorten treatment planning time, and support clinicians with limited experience, thereby enhancing the safety and efficiency of maxillary skeletal expansion in clinical practice.</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

Deep learning–based automated positioning system for maxillary skeletal expander: development and clinical validation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月23日

查看全文
  • Jie Pan,
  • Zhenpeng Liu,
  • Yixin Liu,
  • Yu Zhang,
  • Mengqi Xu,
  • Xiaozhe Li,
  • Xianming Chen,
  • Guangshun Wei,
  • Dongxu Liu

摘要

Objective

This study aimed to develop and validate a deep learning-based artificial intelligence (AI) system capable of automatically generating accurate MSE placement plans to enhance clinical efficiency, safety, and consistency.

Methods

A total of 120 patients with skeletal Class III malocclusion and transverse maxillary deficiency (ages 10–39) were included. Cone-beam computed tomography (CBCT) and intraoral scan data were used to construct individualized three-dimensional anatomical coordinate systems. A deep learning model was trained to automatically segment key anatomical structures (e.g., the incisive foramen and transverse palatine suture) and identify essential craniofacial landmarks. The MSE was then automatically positioned within this coordinate system, while a collision detection algorithm ensured appropriate spacing from the palatal mucosa and avoided vital anatomical structures. Model performance was compared with manual placement using metrics such as mean Intersection over Union (mIoU), Avoidance Success Rate (ASR), Mean Radial Error (MRE), Mean Angular Error (MAE), and Concordance Correlation Coefficient (CCC).

Results

The average mIoU of the segmentation network was 0.75. The ASR exceeded 90%, demonstrating effective anatomical avoidance. The automated system achieved an axial MRE of 0.32 ± 0.32 mm, a three-dimensional (3D) Euclidean distance error of 0.69 ± 0.36 mm, and an MAE of 1.84 ± 2.13°, with CCC values above 0.90 for linear and translational directions and between 0.87 and 0.93 for angular directions. These results showed no statistically significant differences from manual placement (P > 0.05).Bland–Altman analysis further showed minimal mean bias between AI and manual planning (translation bias ≤ 0.047 mm; rotation bias within − 0.26° to 1.52°), with 95% limits of agreement for all parameters remaining within clinically acceptable ranges; a mild proportional bias was observed only for yaw (P = 0.013).The automated process required an average of 3 min, significantly faster than the 45–60 minutes typically required for manual planning.

Conclusion

The proposed deep learning–based automated MSE positioning system enables accurate and efficient MSE digital planning in a retrospective setting, shows promising clinical potential, and may contribute to more standardized and intelligent orthodontic expansion workflows.

Clinical relevance

This deep learning–driven tool may facilitate standardized MSE placement, shorten treatment planning time, and support clinicians with limited experience, thereby enhancing the safety and efficiency of maxillary skeletal expansion in clinical practice.