面向异质化三维口扫大数据的临床可适用语义分割算法研究

本项目聚焦基于深度学习的智能化口腔诊断技术，针对现有三维口内扫描（Intro-oral Scan, IOS）数据语义分割方法错误率高、效率低及临床评估缺乏带来的适用难题，拟突破三维IOS数据语义分割的算法设计和临床评估关键技术挑战，提出临床可适用的语义分割方法。本申请结合深度学习和口腔医学，拟研究：①异质化、高分辨率、大规模三维IOS数据关键特征提取：构建大规模三维IOS数据库，揭示病例数据统计规律，建立关键特征提取方法；②高精度、细粒度、高效率三维IOS数据语义分割算法设计：结合口腔医学知识，建立无监督表征学习模型，构建三维IOS数据语义分割深度神经网络架构和学习范式；③临床可适用三维IOS数据语义分割性能评估机制构建：为提供可信赖的性能分析，构建完备的临床检测评估机制，通过临床测试验证所提算法可适用性。本申请研究成果将为深度学习在智能化口腔诊断系统的交叉应用提供理论和关键技术支撑。

This project focuses on deep learning-enabled intelligent oral diagnosis techniques, paying special attention on the applicability issues caused by the low accuracy, low efficiency and lack of clinical evaluation of the existing semantic segmentation methods for heterogeneous three-dimensional (3D) intra-oral scan (IOS) data. This project aims to solve the main challenges in algorithm design and clinical evaluation of 3D IOS semantic segmentation techniques, and propose the clinically applicable semantic segmentation method. Integrating deep learning techniques with stomatology domain knowledge, this project will: (1) Propose feature extraction methods for heterogeneous, high-resolution and large-scale 3D IOS data. A large-scale public 3D IOS dataset will be built. Based on the statistics of the cases, novel feature extraction methods for 3D IOS data will be proposed. (2) Design algorithms for accurate, fine-grained, and efficient semantic segmentation on 3D IOS data. By combining stomatology domain knowledge, establishing self-supervised representation learning models, and proposing novel neural network architectures and learning paradigms, a novel semantic segmentation framework for 3D IOS data will be established. (3) Develop performance evaluation mechanisms for clinically applicable 3D IOS data semantic segmentation. A comprehensive evaluation mechanism will be developed to provide reliable performance analysis. In order to evaluate and validate the performance gain of the proposed theoretical methods and key techniques, a large-scale clinical test will be carried out. The expected research outcomes of this project will lay the theoretical and technical foundations for the interdisciplinary research and application between deep learning and intelligent oral diagnosis systems.