CASIA OpenIR  > 毕业生  > 硕士学位论文
超声影像下血管及介入器械检测与动态重建
陈玖安
2024-05-16
Pages60
Subtype硕士
Abstract

血管介入手术是当前医学领域中的一项重要技术,它为动脉堵塞、动脉瘤以及其他血管疾病的患者提供了先进的微创诊疗方案。目前X射线仍然是临床上血管介入手术的首选成像手段,它能够清晰地显示血管结构。然而X射线的辐射也为患者及医生带来了潜在的健康风险。为解决这一问题,无辐射的MRI和超声成像技术被视为有效的替代选项。相较于MRI技术,超声成像成本更低且更容易大规模推广。但是为了让超声成像在介入手术中媲美X射线的效果,仍有一系列问题需要得到解决。首先超声成像受限于探头长度,其视野范围不如单张X射线广。其次超声在软组织成像方面表现良好,但对于密度较高的物体如金属介入器械等则效果不佳。再者超声扫查过程中常常会遇到目标成像不稳定的情况。本文旨在提高超声在血管介入手术中的可行性而进行一系列探索,主要内容如下:

首先本文构建了一个新型超声血管介入数据集。数据集由临床上采集的超声图像和视频组成,包含不同成像角度的血管和介入器械类别。通过精确标注影像中每个实例,也为后续各类检测与分割算法的开发提供了一个高质量的训练和验证平台。同时为了克服医疗数据集规模较小的问题,本文设计了契合超声成像特性的数据增强算法。考虑到超声图像中组织的变形,本文采用了一种改进的弹性形变算法以扩展数据规模,确保在对血管组织进行变形的同时,避免对器械产生不必要的形变。同时本文从硅胶体模中采集了一系列模拟穿刺的影像数据,然后应用风格迁移算法进行数据增强,进一步提升了数据集的多样性。

然后本文介绍了基于视频的血管介入超声实例分割网络US-VIS。由于单帧超声图像噪声多且视野狭窄,导致成像过程中常出现目标丢失等情况。本文利用多帧图像的分析来提升对血管和导丝等关键目标的检测精度与连续性。超声视频实例分割网络US-VIS旨在借助历史帧信息辅助对当前帧中目标的进行检测,通过采用对比学习机制实现相邻帧之间信息的深度融合与互相关联。在推理过程中,将历史帧提取到的特征存储到内存库中,然后使用经过训练优化的对比学习网络实现帧间目标的精确匹配,进而有效提高检测的精度及目标追踪连续性。针对介入器械这类小目标的检测,本文设计了专门的优化模块以提升其检测性能。实验表明,超声视频实例分割网络US-VIS保持了与同类模型相媲美的分割精度,同时其在检测连续性方面有显著提升。在对小型目标检测方面,US-VIS在mAP50评价标准下的检测精度相比其他模型提高了3.3%。

最后本文介绍了超声动态三维重建技术。超声图像由于视野限制,无法如其他成像技术那样利用骨骼和皮肤等附加信息进行精确血管定位。动态三维重建技术提供了一种边扫查边重建的方案,以此实时定位血管位置,提高超声诊疗的效率。本文提出了两种超声动态重建的技术,一是基于Marching Cube算法的动态重建技术,它能够在成像平面保持平行的情况下完成超声动态三维重建。二是贝塞尔插值动态重建算法,该算法通过对超声图像序列中的关键点进行插值处理,不仅无需成像平面保持互相平行,而且能在机械臂的协助下高效地生成平滑连续的三维模型。本文着重介绍了贝塞尔插值动态重建过程中的优化算法,为超声三维成像技术的发展和应用提供了新的视角,极大地拓展了其在临床诊断和治疗中的应用范围。

Other Abstract

Vascular intervention is a significant technological advancement in the medical field, offering minimally invasive diagnostic and therapeutic solutions for patients with arterial occlusions, aneurysms, and other vascular disorders. While X-ray imaging remains the preferred modality for guiding such interventions due to its clear visualization of vascular structures, the associated radiation exposure poses potential health risks to both patients and surgeons. To address this issue, radiation-free imaging techniques like MRI and ultrasound have been considered as viable alternatives. Compared to MRI, ultrasound imaging is more cost-effective and easier to adopt on a large scale. However, to rival the efficacy of X-ray in interventional procedures, several challenges with ultrasound imaging must be overcome. Firstly, its field of view is limited by the probe size, making it less extensive than that of a single X-ray image. Secondly, while ultrasound performs well in soft tissue imaging, it is less effective in visualizing high-density objects, such as metallic interventional instruments. Furthermore, target visualization with ultrasound can often be unstable during scanning. This paper aims to enhance the feasibility of ultrasound in vascular interventional surgery, with the following key points:

Initially, this study has developed a novel ultrasound vascular intervention dataset aimed at deepening the understanding of the various instruments used in vascular interventions and the characteristics of vascular ultrasound imaging. Comprising clinically acquired ultrasound images and videos, the dataset includes various imaging angles of vessels and interventional instruments. Precise annotation of each instance in the videos provides a high-quality training and validation platform for the development of detection and segmentation algorithms for various types of ultrasound vessels and instruments. To counteract the issue of small medical dataset sizes, this paper has designed data augmentation algorithms that align with the properties of ultrasound imaging. Considering the deformation of tissues in ultrasound images, an improved elastic deformation algorithm was employed to expand the dataset size while avoiding unnecessary deformation of the instruments. Additionally, a series of simulated puncture imaging data was collected from silicone phantoms, and style transfer algorithms were applied for data augmentation, further enhancing dataset diversity.

Subsequently, the paper introduces the Ultrasound Video Instance Segmentation network (US-VIS) designed for video-based vascular interventional ultrasound. Due to the noisy nature of single-frame ultrasound images and their narrow field of view, which frequently leads to target loss during imaging, the paper leverages multi-frame analysis to improve detection accuracy and continuity for key targets such as vessels and guidewires. US-VIS aims to utilize information from historical frames to assist in detecting targets in the current frame, achieving a deep fusion and interrelation of information between adjacent frames through a contrastive learning mechanism. During inference, features extracted from historical frames are stored in a memory bank, and a trained and optimized contrastive learning network is employed to precisely match targets across frames, thereby effectively enhancing detection accuracy and tracking continuity. For detecting small targets like interventional instruments, specialized optimization modules were designed to boost detection performance. Experiments demonstrate that US-VIS maintains segmentation accuracy comparable to other networks while showing significant improvements in detection continuity. In terms of small target detection, US-VIS exhibited a 3.3% increase in detection accuracy under the mAP50 metric compared to other models.

Finally, the paper discusses ultrasound dynamic 3D reconstruction technology. Due to the limited field of view in ultrasound imaging, which cannot utilize additional anatomical information such as bones and skin for precise vascular localization like other imaging modalities, dynamic 3D reconstruction technology offers a scanning and reconstruction approach that can locate vessels in real-time, thus enhancing the efficiency of ultrasound diagnostics and therapeutics. The study proposes two techniques for ultrasound dynamic reconstruction: one is based on the Marching Cubes algorithm, capable of performing dynamic 3D ultrasound reconstruction while maintaining parallel imaging planes. The other leverages a Bezier dynamic interpolation reconstruction algorithm, which interpolates key points in a series of ultrasound images without the need for imaging plane parallelism and can efficiently generate smooth continuous 3D models with the aid of a robotic arm. The paper focuses on the optimization algorithms used during the Bezier interpolation process, offering new perspectives and methods for the development and application of ultrasound 3D imaging technology, significantly broadening its clinical diagnostic and therapeutic applications.

Keyword血管介入手术,超声成像,数据集构建,视频实例分割,三维动态重建
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/57274
Collection毕业生_硕士学位论文
多模态人工智能系统全国重点实验室_智能微创医疗技术团队
Recommended Citation
GB/T 7714
陈玖安. 超声影像下血管及介入器械检测与动态重建[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
毕业论文04052241.pdf(14629KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[陈玖安]'s Articles
Baidu academic
Similar articles in Baidu academic
[陈玖安]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[陈玖安]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.