CASIA OpenIR  > 毕业生  > 博士学位论文
面向血管介入手术机器人自主导航的视觉感知方法
周彦捷
2022-05-26
页数158
学位类型博士
中文摘要

腹主动脉瘤作为一种常见的退行性血管疾病,一直是严重威胁人类健康的重要原因。近年来,腔内修复术作为一种微创血管介入手术已经逐渐成为腹主动脉瘤的主要治疗方式。但是,随着介入手术需求的迅速增长,手术医师短缺的问题日益严峻。针对这一问题,开发具有自主导航功能的血管介入手术机器人是一个有效的解决方案,其可以将手术医师从介入手术中解放出来。该方案的核心难点在于血管介入手术机器人需要具备手术医师一样的视觉感知能力,具体来说,就是从术中数字减影血管造影(Digital Subtraction Angiography,DSA)和X射线影像中获取实施介入手术所需关键信息的能力。本文在国家自然基金重点项目“血管微创介入手术机器人的基础问题研究”(61533016)等项目的支持下,针对血管介入手术机器人自主导航中的视觉感知这一研究目标,结合手术医师实施介入手术的临床及技术需求,围绕以下四个具体研究内容展开研究:(1)基于DSA影像的腹主动脉瘤形态学分析;(2)基于X射线影像的多器械分割;(3)基于X射线影像的线状器械端点定位和方向检测;(4)基于术中影像的通用分割。本文的主要贡献和创新点如下:

(1)针对基于DSA影像的腹主动脉瘤形态学分析任务,本文提出了一种基于两阶段框架的算法解决方案。其中,第一阶段提出了双分支边界感知网络以实现腹主动脉瘤分割,第二阶段提出了形态学测量算法以实现腹主动脉瘤尺寸测量。在该算法模型中,第一,通过设计边界定位分支,独立对动脉瘤边界进行监督,以提高边界像素的分割精度。第二,在掩码预测分支中,通过设计特征聚合模块来解决动脉瘤几何结构不规则的问题。同时,提出了特征选择模块来消除跳跃连接中不相关的响应和噪声,以进一步提高动脉瘤内部的分割精度。第三,设计了融合模块来加强边界定位分支和掩码预测分支之间的联系。此外,建立了一个腹主动脉瘤的形态学分析数据集。实验结果表明,该算法解决方案能够获得优异的腹主动脉瘤形态学分析性能。

(2)针对基于X射线影像的多器械分割任务,本文提出了一种基于多任务学习框架的深度特征精炼算法。在该算法模型中,第一,设计的特征精炼模块不仅可以捕获跨通道的特征信息,还能够捕获具有方向感知和位置敏感的特征信息,这有助于模型识别我们关注的前景对象。第二,设计的通道校准模块能够重新校准多尺度特征融合的权重,这有助于模型平衡语义信息和细节轮廓信息的重要性。第三,设计的连通性损失函数通过构造连通图,对模型预测的掩码进行连通性的深度监督,有效地改善了导管和导丝分割结果中出现局部断裂的情况。此外,建立了一个多器械的分割数据集。实验结果表明,该算法解决方案能够同时兼顾分割准确性及实时性需求。

(3)针对基于X射线影像的线状器械端点定位和方向检测任务,本文提出了一种基于两阶段框架的算法解决方案。其中,第一阶段提出了深度特征聚合网络以实现线状器械分割,第二阶段提出了位置分析算法以实现线状器械端点定位和方向检测。在该算法模型中,第一,设计的特征注意力模块能够使模型学习选择更具辨别力的特征,以实现准确且稳健的线状器械分割。第二,设计的边界增强模块通过提取低层级的细节轮廓信息,以增强边界的特征表示。第三,提出了支持GPU加速的位置分析算法,能够实时地获取线状器械端点坐标和偏转角度。此外,建立了一个线状器械的端点定位及方向检测数据库。实验结果表明,该算法解决方案能够在确保端点定位及方向检测准确率的同时,进一步提高框架的推理速度,满足手术医师的实时性需求。

(4)针对基于术中影像的通用分割任务,本文提出了一种基于Transformer和卷积神经网络(Convolutional Neural Network,CNN)混合编码的通用分割算法。在该算法模型中,第一,设计了由Shuffle Transformer和ResNet-50组成的混合编码器,该编码器不仅能够通过序列化编码的方式获得全局的上下文信息,而且能够保留精细的局部细节信息。第二,设计的Transformer和CNN融合模块能够消除Transformer分支特征和CNN分支特征之间的语义差异,从而有效地提高局部特征的全局感知能力,并增强全局表示的局部细节。第三,设计了多尺度特征融合的解码器,能够有效地提高模型对不同尺度目标的细节特征描述能力。实验结果表明,该算法解决方案能够在DSA影像和X射线影像的数据集上都获得优异的分割性能。

英文摘要

As a common degenerative vascular disease, abdominal aortic aneurysm has always been an important cause of serious threat to human health. In recent years, endovascular repair as a minimally invasive vascular intervention has become the main treatment for abdominal aortic aneurysms. However, with the rapid growth of the demand for interventional surgery, the shortage of surgeons is becoming increasingly serious. To address this problem, developing a vascular interventional surgery robot with autonomous navigation is an effective solution, which can liberate surgeons from interventional operations. The core difficulty of this solution is that the vascular interventional surgery robot needs to have the same visual perception ability as the surgeon. Specifically, visual perception ability is the ability to obtain key information required for performing interventional surgery from intraoperative Digital Subtraction Angiography (DSA) and X-ray images. Supported by the National Natural Science Foundation of China (61533016) and other projects, this dissertation aims at the research goal of visual perception in robotic navigation of vascular interventional surgery, combined with the clinical and technical requirements of surgeons for interventional surgery, focusing on the following four specific research contents: (1) morphological analysis of abdominal aortic aneurysm based on DSA images; (2) multi-instrument segmentation based on X-ray images; (3) endpoint localization and direction detection of wirelike instruments based on X-ray images; (4) universal segmentation based on intraoperative images. The main contributions and novelties of this dissertation are as follows:

(1) Aiming at the morphological analysis task of abdominal aortic aneurysm based on DSA images, this dissertation proposes a two-stage algorithm. Specifically, in the first stage, a dual-stream boundary-aware network is proposed to achieve abdominal aortic aneurysm segmentation. In the second stage, a morphological measurement algorithm is proposed to achieve abdominal aortic aneurysm size measurement. In this algorithm, first, the aneurysm boundary is independently supervised by designing a boundary localization branch to improve the segmentation accuracy of boundary pixels. Second, in the mask prediction branch, a feature aggregation module is designed to address the issue of irregular aneurysm geometry. Meanwhile, a feature selection module is proposed to eliminate irrelevant responses and noise in skip connections, further improving the segmentation accuracy of the aneurysm interior. Third, a fusion module is designed to strengthen the connection between the boundary localization branch and the mask prediction branch. In addition, a morphological analysis dataset of the aneurysm is established. The experimental results show that the proposed algorithm can obtain excellent morphological analysis performance of abdominal aortic aneurysm.

(2) Aiming at the multi-instrument segmentation task based on X-ray images, this dissertation proposes a deep feature refinement network based on multi-task learning. In this network, first, the designed feature refinement module can capture not only cross-channel feature information, but also direction-aware and position-sensitive feature information, which helps the model identify the foreground object we are concerned with. Second, the designed channel calibration module is able to recalibrate the weights of multi-scale feature fusion, which helps the model balance the importance of semantic information and detailed boundary information. Third, the connectivity loss function is designed to monitor the connectivity of the predicted mask by constructing the connectivity graph, which effectively addresses the issue of local fracture in the segmentation results of the catheter and guidewire. In addition, a multi-instrument segmentation dataset is established. The experimental results demonstrate that the proposed algorithm can take into account both segmentation accuracy and real-time requirements.

(3) Aiming at the task of endpoint localization and direction detection of wirelike instruments based on X-ray images, this dissertation proposes a two-stage algorithm. Specifically, in the first stage, a deep feature aggregation network is proposed to achieve the segmentation of wirelike instruments. In the second stage, a position analysis algorithm is proposed to achieve the endpoint location and direction detection of wirelike instruments. In this algorithm, first, a feature attention module is designed to enable the model to learn to select more discriminative features for accurate and robust wire-like instrument segmentation. Second, a boundary enhancement module is designed to extract low-level detail boundary information to enhance the feature representation of the boundary. Third, a GPU-friendly position analysis algorithm is proposed, which can obtain the endpoint coordinates and deflection angles of wirelike instruments in real-time. In addition, a database of endpoint location and direction detection of wirelike instruments is established. The experimental results indicate that the proposed algorithm can further improve the inference speed of the framework while ensuring the accuracy of endpoint localization and direction detection, and meeting the real-time requirements of surgeons.

(4) Aiming at the universal segmentation task based on intraoperative images, this dissertation proposes a universal segmentation network based on the hybrid encoding of Transformer and convolutional neural network (CNN). In this network, first, a hybrid encoder composed of Shuffle Transformer and ResNet-50 is designed, which can not only obtain global context information through serialization encoding, but also retain fine local detail information. Second, the designed Transformer-CNN fusion module is able to eliminate the semantic difference between Transformer branch features and CNN branch features, thereby effectively improving the global perception ability of local features and enhancing the local details of global representation. Third, a multi-scale feature fusion decoder is designed, which can effectively improve the model's ability to describe the details of objects at different scales. Experimental results show that the proposed algorithm can achieve excellent segmentation performance on both DSA and X-ray image datasets.

关键词腹主动脉瘤 腔内修复术 DSA 影像 X 射线影像 图像分割 形态学 分析 端点定位及方向检测 深度学习 机器人导航
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/48506
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
周彦捷. 面向血管介入手术机器人自主导航的视觉感知方法[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2022.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
博士学位论文_周彦捷.pdf(36889KB)学位论文 限制开放CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[周彦捷]的文章
百度学术
百度学术中相似的文章
[周彦捷]的文章
必应学术
必应学术中相似的文章
[周彦捷]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。