超声医生手法模拟算法与机器人自主扫描关键技术

Ultrasound has the advantages of low price, no radiation and portability, and has become the most widely used diagnosis imaging modality in clinics. The two key factors for accurate ultrasound examination and diagnosis are high-quality images and full-coverage scanning, which depend greatly on the doctor’s operating experience: stable tissue-device contact and full-coverage scanning under physiological breathing and tissue deformation environments, accurate identification and measurement of suspected lesions, and a high degree of coordination between doctors' hands, eyes and brain in limited time, these problems have seriously hindered the standardization of ultrasound examination and the consistency of diagnosis. This project aims at solving the key scientific issues including ultrasound expert scanning experience transfer-reinforcement learning convergence, quantitative mathematical modeling of ultrasound image quality, generalization of feature extraction in ultrasound video self-supervised learning and low transient distortion of ultrasound scan images in dynamic environments. The research focuses on the characterization and learning of ultrasound experts’ operating experience, mathematical modeling for quantitative evaluation of ultrasound image quality, environmental status recognition based on multi-source video information, autonomous ultrasound scan planning and image quality control in dynamic environments, and system integration and clinical experiment verification will be conducted. This project will form the basic theory and common technology system of autonomous ultrasound robots, and lay the foundation for promoting the clinical application of safe and reliable medical robots.

超声检查具有价格低、无辐射、可便携等优点，已成为目前临床应用最为广泛的影像诊断方式。超声精确检查和诊断的二大关键因素是高质量的图像和全覆盖的扫描，这极大地依赖医生的操作经验：生理呼吸及组织变形环境下组织-器械稳定的接触力和全覆盖扫描、对疑似病灶的精确识别和测量、有限时间下医生手-眼-脑的高度协调和配合，当前这些问题严重阻碍了超声检查的规范化和诊断的一致性。本项目针对基于先验知识的超声专家扫描经验迁移-强化学习收敛性、超声图像质量的量化数学建模、超声视频自监督学习中特征提取的普适性及动态环境下超声扫描图像的低瞬态失真等关键科学问题，重点研究超声专家操作经验表征与学习、超声图像质量的量化评价数学模型、基于多源视频信息的环境状态识别、动态环境下的超声自主扫描规划与图像质量控制等，并开展系统集成及临床实验验证，形成自主超声机器人基础理论和共性技术体系，为推动安全、可靠的医疗机器人走向临床奠定基础。

超声检查具有价格低、无辐射、可便携等优点，已成为目前临床应用最为广泛的影像诊断方式。超声精确检查和诊断的二大关键因素是高质量的图像和全覆盖的扫描，这极大地依赖医生的操作经验：生理呼吸及组织变形环境下组织-器械稳定的接触力和全覆盖扫描、对疑似病灶的精确识别和测量、有限时间下医生手-眼-脑的高度协调和配合，当前这些问题严重阻碍了超声检查的规范化和诊断的一致性。本项目在国家自然科学基金的支持下，首先建立了临床专家高质量超声图像扫描的操作经验模型，提出了基于强化学习的超声自主扫描机器人经验学习方法，极大提高了超声自主扫描的效率和质量；研究了超声图像质量的量化评价数学模型，提出了基于深度学习的超声全局图像质量评价方法和基于病灶分割的半参考局部超声图像质量评价方法，为超声自主扫描机器人稳定获取高质量超声图像奠定了基础；研究了超声图像/视频的病变智能检测，提出了基于迁移学习的超声图像病灶分割方法和基于动态并行时空Transformer的乳腺超声视频病变分割方法，有效提高了高通量超声图像病变检测的准确性和规范性；研究了动态环境下超声机器人自主扫描规划和图像质量控制，提出了基于多角度点云建模的目标区域均匀全覆盖扫描路径自动规划方法和基于探头-组织交互建模的力-速度混合控制方法，有效提高超声机器人自主扫描的准确性和安全性。集成超声成像系统、力/扭矩传感器、RGB-D相机、乳腺模型等，搭建了超声自主扫描机器人系统，通过乳腺体模扫描模型实验、超声机器人引导穿刺动物实验、人体甲状腺扫描临床实验，对项目所提出的超声自主扫描机器人系统与关键技术进行了临床适用性与安全性验证。发表或录用SCI/EI 学术论文20篇，其中SCI 论文16篇、EI 论文4 篇；共申请发明专利16 项，其中国际PCT专利3项。项目成果形成了自主超声机器人基础理论和共性技术体系，有力推动了安全、可靠的医疗机器人走向临床。