摘要
针对基于深度强化学习的自主超声扫描方法存在训练扫描精度低、训练时间长、扫描任务成功率较低的问题,提出了一种基于改进型多模态信息融合深度强化学习的自主超声扫描方法。首先,该方法融合了超声图像、双视角探头操作图像和6D触觉反馈提供全面的多模态感知信息。为精准捕捉多模态中的时空信息和实现多模态特征的高效融合,设计了一个基于自注意力机制(self-attention mechanism,SA)的多模态特征提取与融合模块。其次,将机器人的6D位姿动作决策任务建模为深度强化学习问题。为贴近专业超声从业医生的操作,设计了混合奖励函数。最后,为解决深度强化学习训练中出现的局部最优和收敛速度慢的问题,提出了DSAC-PERDP(discrete soft actor-critic with prioritized experience replay based on dynamic priority)算法。在真实环境中的测试表明,该方法在扫描精度、任务成功率和训练速度方面较基线模型分别提升了49.8%、13.4%和260.0%,在干扰条件下仍保持良好性能。实验证明,该方法显著提升了扫描精度、任务成功率和训练速度,并具有一定的抗干扰能力。
To address the issues of low training accuracy,prolonged training time,and low success rate of scanning tasks in ultrasound scanning based on deep reinforcement learning(DRL),this paper proposed an autonomous ultrasound scanning method based on improved multimodal information fusion and DRL.Firstly,the method integrated ultrasound images,dual-view probe manipulation images,and 6D tactile feedback to provide comprehensive multimodal perception.To accurately capture spatiotemporal information in multimodal data and achieve efficient feature fusion,this paper designed a multimodal feature extraction and fusion module based on the self-attention mechanism(SA).Secondly,it formulated the 6D pose decision-making task for the robot as a DRL problem.And this paper designed a hybrid reward function to emulate to professional ultrasonographers.Lastly,to address local optima and slow convergence in DRL training,this paper introduced the DSAC-PERDP algorithm.Tests in real environments demonstrate that the proposed method improves scanning accuracy,task success rate,and training speed by 49.8%,13.4%and 260.0%,respectively,compared to baseline models.Moreover,the method maintains robust performance under interference conditions.These findings validate that the proposed approach not only signi-ficantly improves scanning accuracy,task success rate,and training efficiency but also exhibits notable anti-interference cap-abilities.
作者
徐加开
陆奇
李祥云
李康
Xu Jiakai;Lu Qi;Li Xiangyun;Li Kang(College of Electrical Engineering,West China Hospital,Sichuan University,Chengdu 610065,China;Sichuan University-Pittsburgh Institute,West China Hospital,Sichuan University,Chengdu 610065,China;West China Biomedical Big Data Center,West China Hospital,Sichuan University,Chengdu 610065,China)
出处
《计算机应用研究》
北大核心
2025年第6期1624-1631,共8页
Application Research of Computers
基金
国家自然科学基金资助项目(51805449,62103291)
四川省科技计划资助项目(2024YFFK0033,2023YFH0037,2023ZHCG0075,2023YFG0057,2022YFS0021,2022YFH0073)
四川大学华西医院医工交叉融合人才培养基金资助项目
四川大学华西医院1·3·5卓越学科项目(ZYYC21004,ZYJC21081)。
