毕业生知识库

      颅内动脉粥样硬化狭窄（Intracranial atherosclerotic stenosis，ICAS）是引发缺血性脑卒中的主要原因之一，在我国卒中人群中的发病率高达33% ~50%。当颅内动脉狭窄程度（横截面上斑块面积/血管面积）超过70% ，且出现明显的脑部缺血性症状时，需要通过血管内介入手术进行治疗。在颅内动脉介入手术中，需要先将导丝依次经过主动脉弓、颈总动脉、颈内动脉颅外段，准确递送至颅内动脉病变位置，建立治疗通道，才能递送后续的治疗器械进行治疗。然而，由于颅内动脉和颈内动脉颅外段（本文中统称为脑血管）血管走形迂曲，病变情况复杂，术中单一的2D-DSA 影像往往无法反映血管完整的三维结构信息，术者在手术过程中仅能凭借有限角度的二维投影和自身经验对导丝进行操作，这容易导致导丝尖端在操作过程中对血管壁造成损伤，甚至刺穿血管壁，引发严重的手术并发症。

       针对上述挑战，开发基于多模态影像信息融合的颅内动脉介入手术导航系统是一种有效的解决方案。该方案的核心思路为：首先对2D-DSA 影像中的脑血管进行精准分割，并对三维影像中的脑血管进行快速精准分割；而后，将不同模态三维影像中反映的血管整体走形（TOF-MRA）、缺失血管段走形（HR-MRI）、真实管腔形态（3D-DSA）等信息进行融合，获得血管完整的三维结构，并从中提取血管形态学信息；最后，根据2D-DSA 影像和3D-DSA 影像中的血管走形，将完整的三维血管结构配准到病人的术中位姿上，实现三维动态导航，提高手术的安全性。本文在国家自然科学基金面上项目“面向复杂病变的血管介入手术机器人自主操控关键技术研究”（62073325）、青年科学基金项目“基于多模态信息融合的血管介入手术机器人三维智能引导与安全操控研究”（62303463）、国家重大仪器研制项目“面向结构性心脏病的多维智能超声引导精准介入诊疗系统的研制”（82327801）等项目的支持下，围绕在颅内动脉介入手术中实现基于多模态医学影像信息融合的三维动态导航这一总体目标，对其中的关键技术进行了研究与探索。本文的主要工作和贡献可以归纳如下：

（1）基于管状特征提取和全局注意力编码的脑血管2D-DSA 影像的精准提取方法。针对2D-DSA 影像中颅内动脉末端血管结构纤细、重叠严重与血管整体形态差异大等问题，本文提出了一种基于管状特征提取与全局注意力编码的二维血管分割模型，该模型兼顾了对局部管状结构信息和全局血管形态信息的编码，实现了对2D-DSA 影像中脑血管的精准分割。该网络模型在本文构建的IAS-2D-DSA 数据集（包含331 例脑血管2D-DSA 影像）上取得了最佳的分割精度（Dice 指标均值为84.01%，mIoU 指标均值为74.37%，clDice 指标均值为88.60%，HD95 指标均值为16.70）。

（2）基于全局特征融合与动态窗口采样的脑血管3D-DSA 影像的快速提取方法。针对脑血管3D-DSA 影像分割任务中全局语义缺失与分割时间较长的问题，本文依次提出了一种基于全局信息融合的双分支三维血管分割网络（Region-Global Fusion Net，RGFNet）和一种基于动态窗口采样的快速三维血管分割方法（Fast 3D Vessel Segmentation method，FVS-3D）。其中，RGFNet 通过引入全局分割分支对影像的全局依赖信息进行逐级提取，并通过融合模块将其补充到局部分割分支中，提高了模型的分割精度。FVS-3D 方法通过对血管待分割区域进行动态自适应的采样来减少待分割窗口的个数，从而降低模型推理时间，并采用了基于解剖依赖的空间融合模块对血管的全局解剖信息进行捕获，从而提高模型的分割精度。RGFNet 和FVS-3D 分割方法分别在IAS-3D-DSA 数据集（包含210 例脑血管3D-DSA 影像）上取得了最佳的分割精度（Dice 指标均值为93.36%，mIoU 指标均值为87.83%，HD95 指标均值为15.64）和实现了模型推理时间的大幅度降低（在IAS-3D-DSA 数据集与SegTHOR 公开数据集上的CPU 推理时分别平均降低了99.43%）。

（3）多模态影像融合导航系统集成。针对单模态影像提供的血管介入手术引导信息不全面的问题，本文基于得到的脑血管二维影像和三维影像的分割结果，提出了一种多模态医学影像信息融合的技术路线，实现了多模态影像导航系统的集成。首先，本文通过一种基于中心线点匹配的3D/3D 血管影像配准方法，对TOF-MRA 影像、HR-MRI 影像和3D-DSA 影像中反映的不同的血管结构进行融合，从而获得了完整的血管空间走形与真实的管腔形态结构，并随之对融合影像中的形态学参数进行自动提取。而后，本文提出了一种基于双投影匹配的2D/3D血管影像配准方法，从而将脑血管的术前三维融合影像配准到病人的术中位姿上，最终实现了三维动态导航，能够为术者提供多维度、多视角的手术引导信息，降低颅内动脉介入手术的操作难度与风险系数。

Intracranial atherosclerotic stenosis (ICAS) is one of the main causes of ischemic stroke, with the incidence rate among the stroke population in my country being as high as 33% and 50%. When the degree of intracranial artery stenosis (plaque area/vessel area in cross section) exceeds 70% and obvious symptoms of cerebral ischemia occur, endovascular interventional surgery is required. In intracranial artery interventional surgery, the guidewire needs to be passed through the aortic arch, common carotid artery, and extracranial segment of the internal carotid artery in sequence, and accurately delivered to the location of the intracranial artery lesion, so that a channel can be established before subsequent instruments can be delivered for treatment. However, due to the tortuous nature of the intracranial arteries and the extracranial segments of the internal carotid arteries (collectively referred to as intracranial and extracranial arteries in this article), and the complex lesions, a single 2D-DSA image during surgery is not able to reflect the complete 3D structure of the vessels. The operator can only rely on the 2D projection of limited angles and his or her own experience to deliver the guidewire during the surgery, which can easily cause the tip of the guidewire to damage the vessel wall during delivery, or even pierce the vessel wall, causing serious surgical complications.

To overcome the challenges mentioned above, developing an intracranial artery interventional surgery navigation system based on multi-modal image information fusion is an effective solution. The core of this solution is: Firstly, the cerebral vessels in the 2D-DSA and the 3D images is segmented quickly and accurately; Then, the complete 3D structure of the vessel is obtained by fusing the information such as the overall shape of the vessels (TOF-MRA), the shape of the missing vessel segment (HR-MRI), and the true lumen shape (3D-DSA), and extracting morphological information from the fused image; finally, based on the vessel shape in the 2D-DSA image and 3D-DSA image, the complete 3D vessel structure is registered to the intraoperative posture to achieve 3D dynamic navigation. Supported by the Youth Science Fund project ”Three-dimensional intelligent guidance and safe control of vascular interventional surgery robots based on multi-modal information fusion” (62303463), the National Natural Science Foundation of China project ”Research on key technologies for autonomous control of vascular interventional surgery robots for complex lesions” (62073325), and the National Major Instrument Development project ”Development of a multi-dimensional intelligent ultrasound-guided precision interventional diagnosis and treatment system for structural heart disease” (82327801), this thesis focused on the goal of realizing 3D dynamic navigation fo intracranial artery interventional surgery based on multi-modal medical image information fusion, and explored the key methods for it. The main contributions of this thesis can be summarized as follows:

(1) Accurate segmentation method of cerebral vessels in 2D-DSA based on tubular feature extraction and global attention encoding. Aiming at the problem of limited segmentation accuracy in 2D-DSA images of cerebral vessels due to slender vessel structure and serious overlapping in the end area of intracranial artery, and the different vessel morphology, a 2D vessel segmentation network based on the tubular features extraction and global attention encoding is proposed. The proposed segmentation model takes account of both the capturing of local tubular structure information and encoding of global vessel morphology information, thereby achieving accurate segmentation of cerebral vessels in 2D-DSA images. This segmentation model achieved the best segmentation accuracy on the IAS-2D-DSA dataset (Dice 84.01%, mIoU 74.37%, clDice 88.60%, HD95 16.70).

(2) Fast segmentation method of cerebral vessels in 3D-DSA based on global feature fusion and dynamic window sampling. Aiming at the problems of the lack of global semantic and long inference time in the 3D-DSA image segmentation task of intracranial and extracranial arteries, this thesis proposes Region-Global Fusion Net (RGFNet) based on global information fusion and a fast 3D Vessel Segmentation method (FVS-3D) based on dynamic window sampling. RGFNet introduces the global segmentation branch to extract the global dependency of the image step by step, and supplements it into the local segmentation branch through the fusion module, which improves the segmentation accuracy of the model. FVS-3D reduces the number of windows to be segmented by dynamically sampling of the vascular region, thereby reducing model inference time, and an anatomical-dependence-based spatial fusion module is proposed to capture the global anatomical information of vessels, thereby improving the segmentation accuracy of the model. RGFNet and FVS-3D segmentation methods respectively achieved the best segmentation accuracy (Dice 93.36%, mIoU 87.83%, HD95 15.64) on the IAS-3D-DSA dataset (including 210 cerebral vessels 3D-DSA images) and a substantial reduction in model inference time (an average 99.43% reduction in CPU inference time is achieved on both the IAS-3D-DSA dataset and the SegTHOR public dataset).

(3) The integration of multi-modal image fusion navigation system. In order to solve the problem of incomplete guidance information for vascular interventional surgery provided by single-modal images, this thesis proposes a multi-modal medical image information fusion method based on the segmentation results of the obtained 2D and 3D cerebrovascular images, realizing the integration of multi-modal image navigation systems. First, this paper uses a 3D/3D vascular image registration method based on skeleton point matching to fuse different vascular information reflected in TOF-MRA, HR-MRI and 3D-DSA, thereby obtaining a complete vessel structure and real lumen morphology; Then, this thesis proposes a 3D morphological analysis algorithm to automatically extract the key morphological parameters of cerebral vessels in the fused image; Finally, this paper proposes a 2D/3D vascular image registration method based on dual projection matching, thereby registering the preoperative 3D fused image of cerebral vessels to the patient’s intraoperative posture, realizing 3D dynamic navigation, providing The operator multi-dimensional and multi-perspective surgical guidance information to reduce the difficulty and risk of intracranial artery interventional surgery.