基于深度学习的牙齿预备体倒凹检测研究

Research on Undercut Detection in Dental Preparations Based on Deep Learning

摘要: 在进行牙体缺损修复治疗时，预备体倒凹区域的准确检测对于后续修复体设计工作具有至关重要的意义。尽管已有一些研究尝试解决这一问题，但它们在实际应用中通常面临着效率低下和精度不高的挑战，无法为牙科医生提供更加准确和直观的意见。为此本文提出了一种基于自监督的Dental-UDF网络和OBBTree线面碰撞结合的牙齿预备体倒凹检测方法。首先使用结合场一致性优化与零水平集UDF约束的混合优化方法从预备体点云中准确学习无符号距离场，再使用Marching Cubes算法提取连续的网格曲面。然后为网格模型构建OBBTree，通过批量平行线投射方式准确检测倒凹区域。实验结果显示，结合场一致性优化与零水平集UDF约束的网络重建模型在各项评价指标中都达到了最优，线面碰撞后的倒凹百分比误差介于0.215%至0.544%之间。这些结果验证了本文方法在自动化倒凹检测方面的高效性和准确性，为提高牙齿修复工作的质量和效率提供了有力的技术支持。

Abstract: In the treatment of dental defects, the accurate detection of undercut areas in the preparation is of critical importance for the subsequent design of the restoration. Although some studies have attempted to solve this issue, they often face challenges of low efficiency and inaccuracy in practical applications, failing to provide dentists with more precise and intuitive advice. Therefore, this paper proposes a tooth preparation undercut detection method that combines a self-supervised Dental-UDF network with OBBTree line-surface collision. Firstly, a hybrid optimization method that integrates field consistency optimization with zero-level set UDF constraints is used to accurately learn the unsigned distance field from the preparation point cloud, followed by the use of the Marching Cubes algorithm to extract continuous mesh surfaces. Then, an OBBTree is constructed for the mesh model, and the undercut area is accurately detected through batch parallel line projection. Experimental results show that the network reconstruction model combining field consistency optimization with zero-level set UDF constraints achieved optimal performance in all evaluation metrics, with the undercut percentage error ranging between 0.215% and 0.544%. These results verify the method’s efficiency and accuracy in automatic undercut detection, providing strong technical support for improving the quality and efficiency of dental restoration work.