Related Experiment Video
Updated: Sep 4, 2026
Reliability of Artificial Intelligence-Based Cone Beam Computed Tomography Integration with Digital Dental Images
Published on: February 23, 2024
Artificial intelligence-enabled automatic segmentation of impacted mandibular third molars: A comprehensive
Sum Yin Au Yeung1, Wai Ying Kot1, Yin Yan Leung1
1Faculty of Dentistry, The University of Hong Kong, 34 Hospital Road, 999077, Hong Kong SAR, China.
Background/Purpose:
Impacted mandibular third molar models generated from cone-beam computed tomography images facilitates preoperative assessment and surgical planning. Despite recent applications of artificial intelligence (AI) in tooth segmentation, its performance in mandibular third molar segmentation remains underexplored. This study aims to develop a comprehensive comparison of available algorithms for mandibular third molar segmentation for clinical applications and research.
Materials And Methods:
Forty impacted mandibular third molars were segmented using interactive thresholding (Thresholding), Blue Sky Plan (BSP AI), Blue Sky Plan with manual adjustments (BSP AI with adjustment), and Planmeca Romexis (Romexis AI), with manual segmentation as reference standard. Segmentation accuracy was assessed using dice similarity coefficient (DSC), intersection over union (IoU), average symmetric surface distance (ASSD), 95 % Hausdorff distance (95HD) and relative volume difference (RVD). Segmentation quality was evaluated based on morphology, three-dimensional visualization, and spatial relationships with adjacent structures.
Results:
BSP AI with adjustment achieved the highest accuracy (DSC: 0.89 ± 0.02, 95HD: 0.85 ± 0.25 mm), followed by Thresholding (DSC: 0.86 ± 0.03; 95HD: 1.28 ± 0.23 mm) and BSP AI (DSC: 0.85 ± 0.08; 95HD: 1.91 ± 1.91 mm). Romexis AI (DSC: 0.81 ± 0.03; 95HD: 1.41 ± 0.32 mm) demonstrated the poorest performance. BSP AI with adjustment achieved the highest quality score of 38.00 (38.00-39.00), similar to Thresholding achieving 38.00 (37.25-39.00), followed by BSP AI and Romexis AI being the last.
Conclusion:
AI-enabled segmentation achieved high accuracy and efficiency in mandibular third molar segmentation, while manual adjustment ensured optimal quality. Further research is warranted to promote future enhancement and clinical applications of AI-enabled mandibular third molar segmentation.
