姿态估计模型用于识别主动脉腔内介入器具  

Pose estimation model for recognizing instruments in intraluminal aortic intervention

作　　者：伍尚至 陆清声 WU Shangzhi;LU Qingsheng(School of Health Science and Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Department of Vascular Surgery,the First Affiliated Hospital ofNaval Medical University,Shanghai 200433,China)

机构地区：[1]上海理工大学健康科学与工程学院,上海200093 [2]海军军医大学第一附属医院血管外科,上海200433

出　　处：《中国介入影像与治疗学》2025年第2期136-141,共6页Chinese Journal of Interventional Imaging and Therapy

基　　金：科技创新2030重大项目“新一代人工智能”专项(2018AAA0102603);上海市科技计划项目(23XD1405000)。

摘　　要：目的建立姿态估计模型,观察其识别主动脉腔内介入器具的性能及临床适用性。方法收集45例接受主动脉介入治疗(包括单纯医师及医师在机器人辅助下介入操作)患者的48段数字减影血管造影(DSA)视频片段并经预处理拼接为1个视频、含10220帧图像,根据图像背景复杂度将其分为A~E共5种。于其中隔帧抽取5110帧作为关键数据集,并与其内划分训练集(n=4599)与验证集(n=511),以人工标注训练集各帧内所有器具,包括导丝、导管、导丝尖端及导管尖端并加以累计。利用模型主干网络提取特征,结合光流跟踪算法实现关键点邻帧跟踪;针对10220帧图像进行逐帧推理后,于训练集统计模型分类混淆性能,于验证集计算模型精度,于全部数据集获取推理速度并与U-Net进行对比。结果所获模型识别验证集导丝及导管的平均精度分别为0.832、0.808,均优于U-Net的0.767、0.793;其识别医师操作下验证集胸主动脉内导丝、导管的平均精度为0.787、0.756,识别腹主动脉内导丝、导管的平均精度分别为0.826、0.806;识别机器人辅助操作下导丝、导管的平均精度为0.855、0.834。A~E 5种背景下,模型识别验证集导丝精度分别为0.594、0.865、0.817、0.793及0.764,识别导管精度分别为0.626、0.847、0.795、0.739及0.694。模型累计于训练集正确分类7971、7026、7551及7533个实例的导丝、导管、导丝尖端及导管尖端,累计误分类1271、1357、812及863个实例的导丝、导管、导丝尖端及导管尖端。针对全部10220帧,模型平均推理帧频为20.66帧/秒。结论所获姿态估计模型识别血管腔内导丝、导管精度和实时处理能力出色,可用于临床。Objective To establish a pose estimation model for recognizing instruments in intraluminal aortic intervention and to evaluate its efficiency as well as clinical applicability.Methods Forty-eight digital subtraction angiography(DSA)video clips of 45 patients who underwent intraluminal aortic intervention by physician or robot-assisted physician were collected and preprocessed into one video with totally 10220 frames.The frames were categorized into 5 levels of background complexity.A key dataset of 5110 frames was selected from 10220 frames after intermittent frame extraction and divided into training set(n=4599)and validation set(n=511).The accumulative numbers of guidewires,catheters,guidewire tips and catheter tips with manual labels in training set were recorded.Feature extraction was performed via a backbone network of this model,and an optical flow tracking algorithm was combined to achieve key point tracking between consecutive frames.After frame-by-frame inference across all 10220 frames,the confusion results in training set were calculated,while the accuracy was calculated in validation set,the inference speeds were record in all sets,and the results were compared with those of U-Net.Results In validation set,the mean accuracy of the model for recognizing guidewires and catheters was 0.832 and 0.808,respectively,better than those of U-Net(0.767 and 0.793).For physician-operated procedures,the average accuracy of this model for recognizing guidewires and catheters within chest aorta was 0.787 and 0.756,which within abdominal aortic was 0.826 and 0.806,respectively.For robot-assisted procedures,the average accuracy of this model for recognizing guidewires and catheters in validation set was 0.855 and 0.834,respectively.Under 5 descending levels of background complexity,the model achieved recognizing accuracies of 0.594,0.865,0.817,0.793 and 0.764 for guidewires,and of 0.626,0.847,0.795,0.739 and 0.694 for catheters.In training set,this model accumulatively correctly classified guidewires,catheters,guidewir

关 键 词：血管内操作 放射摄影术 深度学习 模式识别 视觉 

分 类 号：R543.5[医药卫生—心血管疾病] R445.4[医药卫生—内科学]