基于注意力机制的手术器械多任务追踪方法  

Multi-task tracking method for surgical instruments based on attention mechanism

作　　者：蔡琼[1] 王浩[2] 柯水洲 CAI Qiong;WANG Hao;KE Shuizhou(Information Construction and Development Center, Hefei University of Technology, Hefei 230009, China;School of Management, Hefei University of Technology, Hefei 230009, China)

机构地区：[1]合肥工业大学信息化建设发展中心,安徽合肥230009 [2]合肥工业大学管理学院,安徽合肥230009

出　　处：《合肥工业大学学报（自然科学版）》2022年第6期746-752,共7页Journal of Hefei University of Technology：Natural Science

基　　金：国家自然科学基金重大研究计划培育资助项目(91846107)。

摘　　要：由于微创手术视野小,同时手术过程中存在多次器械置换和位置矫正,导致手术过程中存在着器械置换错误,甚至是器械误伤组织的风险。针对这一问题,文章提出一个基于改进RCNN(Region-based Convolutional Neural Network)的微创手术多任务器械追踪方法以追踪手术器械的使用情况。首先将注意力机制引入RPN(Region Proposal Network)中,利用具有注意力机制的注意力模块,充分提取手术器械末端执行器的特征,使得模型可以在ResNet-101网络中取得更快的训练收敛速度;其次改进了目标损失函数,使得模型在目标追踪的同时可以进一步获取手术器械的轮廓,实现手术器械多任务追踪。该文在真实的临床数据中进行了模型验证,实验结果表明:改进的RCNN模型在训练过程相较于ResNet-50模型有更快的收敛速度;同时在单个任务时的准确度、召回率和平均精度均值指标分别为66.4%、59.5%和39.2%,比基于ResNet-101的RCNN模型分别平均提高2%。Due to the limited field of view of minimally invasive surgery and multiple instrument replacements and position corrections during the surgery,the patient faces a risk of instrument replacement errors and even tissue injury by surgical instruments during surgery.To address this problem,this paper proposes an improved Region-based Convolutional Neural Network(RCNN)based multi-task instrument tracking method for minimally invasive surgery to track the surgical instruments.First,the attention model is applied to the Region Proposal Network(RPN)of the model.It will fully extract the features of the end effector of the surgical instrument,which helps the model achieve faster training convergence in the ResNet-101 network.Then the loss function is improved,enabling the model to obtain the contour of the surgical instrument further while tracking the target.The improved model has been evaluated on the real clinical dataset.Experimental results show that the improved RCNN model has a faster convergence rate than the ResNet-50 model during the training.The model achieves the precision,recall,and mAP with 66.4%,59.5%,and 39.2%,respectively,averagely improved 2%than the ResNet-101 based RCNN model.

关 键 词：注意力机制 计算机辅助决策 目标检测 多任务算法 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]