Collateral circulation assessment in acute ischemic stroke: current applications and future directions of deep learning
-
摘要:
在急性缺血性脑卒中患者中，侧支循环对维持缺血半暗带灌注、延缓梗死进展及改善血管内治疗结局具有关键作用。目前，侧支循环评估多借助CT血管成像、多时相CT血管成像、CT灌注成像、磁共振灌注成像或数字减影血管造影，仍主要依赖Tan评分、Maas评分及美国介入和治疗神经放射学会/介入放射学会侧支分级等人工评分工具，存在主观性强、可重复性不足等问题，难以满足临床快速、客观、量化的评估需求。在再灌注治疗精细化的背景下，准确评估侧支代偿能力有助于解释梗死进展的异质性及无效再通现象（即血管再通但组织未实现有效再灌注）。近年来，深度学习技术的发展为侧支循环的自动化评估和精细化量化提供了新的技术路径，相关研究涵盖侧支循环自动分级、脑血管结构提取与多尺度量化分析，以及融合灌注参数与临床信息的多模态预测模型等多个方面。尽管现有研究已取得一定进展，但仍面临数据稀缺、样本类别不平衡、数据域偏移显著及缺乏统一分级标准等诸多挑战。本文综述了深度学习在急性缺血性脑卒中侧支循环评估中应用的研究进展与关键瓶颈，并对多中心数据标准化、动态血流建模、自监督学习、可解释人工智能及人工智能与临床诊疗流程融合等未来发展方向进行展望，旨在推动构建更稳定、可推广的侧支循环评估框架。
Abstract:In acute ischemic stroke, collateral circulation plays a critical role in maintaining ischemic penumbral perfusion, delaying infarct progression, and improving outcomes after endovascular treatment (EVT). Currently, collateral assessment mostly relies on computed tomography angiography (CTA), multiphase CTA, computed tomography perfusion, magnetic resonance perfusion, or digital subtraction angiography, and still mainly depends on visual grading systems such as the Tan, Maas, and American Society of Interventional and Therapeutic Neuroradiology/Society of Interventional Radiology scores. These approaches are subjective and show limited reproducibility, failing to meet the clinical demand for rapid and objective quantitative evaluation. In the context of increasingly refined reperfusion strategies, accurate evaluation of collateral capacity is essential for explaining heterogeneity in infarct progression and the futile recanalization, defined as successful recanalization without effective tissue reperfusion. In recent years, deep learning methods have been applied to enable automated quantitative assessment of collateral circulation. Existing studies have focused on automated collateral grading, extraction and multiscale quantification of vascular structure, and multimodal predictive models integrating perfusion parameters with clinical information. Despite encouraging progress, challenges remain, including limited data availability, class imbalance, domain shift, and the absence of unified grading standards. This review summarizes recent advances and key bottlenecks of deep learning in the assessment of collateral circulation in acute ischemic stroke and discusses future directions, including multicenter standardization, dynamic blood flow modeling, self-supervised learning, explainable artificial intelligence, and integration into clinical workflows, so as to facilitate more robust and generalizable collateral assessment frameworks.
-
急性缺血性脑卒中（acute ischemic stroke，AIS）是全球范围内致残和致死的主要病因之一，侧支循环在急性缺血过程中的保护性作用已被大量研究证实[1]。充足的侧支灌注可有效维持缺血半暗带血流、延缓梗死核心灶进展，并显著改善血管内治疗的功能预后；反之，侧支代偿不足则常伴随脑组织缺血快速恶化、再灌注损伤风险增高以及临床预后不良[2]。侧支循环在脑血流动力学调控、脑组织命运决定和治疗反应预测中的核心作用，使其成为AIS评估体系中不可或缺的关键环节。
从生理角度而言，侧支循环并非单一维度概念，可从动脉层面、组织灌注层面及静脉层面进行多维度综合解析。动脉层面的侧支代偿状态主要反映脑动脉闭塞远端脑部血管的代偿性充盈情况，评估通常基于计算机断层扫描血管成像（computed tomography angiography，CTA）、多时相CTA（multiphase CTA，mCTA）或数字减影血管造影（digital subtraction angiography，DSA），采用Tan评分、Maas评分及美国介入和治疗神经放射学会/介入放射学会（American Society of Interventional and Therapeutic Neuroradiology/Society of Interventional Radiology，ASITN/SIR）侧支分级等视觉评分体系[3]。组织灌注层面的侧支评估则聚焦于侧支血流在脑微循环中的灌注效能，主要依赖计算机断层扫描灌注成像（computed tomography perfusion，CTP）的量化参数指标，如低灌注强度比（hypoperfusion intensity ratio，HIR），间接反映脑组织对缺血的耐受性[4]。近年来，静脉层面的侧支评估逐渐受到临床关注，通过分析脑静脉充盈程度与对称性，如皮质静脉显影评分，评价整体脑血流量及侧支灌注能否形成有效的代偿血流回路[5]。
尽管上述多层级评估方法从不同维度刻画了侧支循环的关键特征，但当前临床实践中仍以人工视觉判读为主，整体属于定性或半定量评估框架，且各层级指标常被割裂使用。这一评估模式易受扫描协议差异、时间窗设置、造影剂动力学特性以及阅片者临床经验等因素影响，导致不同中心、不同设备间的评估结果一致性有限[6]。随着再灌注治疗策略的不断精细化，传统侧支评估方法在评估速度、客观性及可重复性方面的局限性愈发凸显，难以满足AIS精准化管理的临床需求。
近年来，深度学习技术在医学影像分析领域的快速发展，为侧支代偿的自动化评估与多层级量化提供了新的技术路径。通过对CTA、mCTA、非增强CT（non-contrast CT，NCCT）、CTP或动态磁敏感对比增强磁共振成像（dynamic susceptibility contrast magnetic resonance imaging，DSC-MRI）等影像中的血管结构特征、灌注动力学信息和跨模态语义关系进行建模，深度学习能够捕捉到人类视觉难以识别的影像纹理差异与隐含灌注线索，进而提升侧支评估的一致性、精准度与可解释性[7]。其优势在于可实现端到端的影像理解，使侧支代偿状态的判定更具客观性，同时具备融入临床诊疗决策流程的潜在价值[8]。
在此背景下，系统梳理深度学习在侧支代偿量化评估领域的研究现状与发展趋势尤为必要。本文围绕当前该领域的核心技术任务——侧支循环自动分级、脑血管结构提取与量化、多模态信息融合及组织命运预测——全面总结现有研究的模型设计思路、数据来源类型、模型性能表现及临床适用场景。同时，本文探讨了该领域面临的关键挑战，如研究数据稀缺、标签标注不一致、跨中心模型泛化能力不足、灌注动力学建模受限等，并展望其在动态影像融合、自监督学习、可解释人工智能及临床诊疗流程整合等未来发展方向的潜在突破路径。
1 侧支循环的病理生理机制及其影像学表征
侧支循环在AIS中发挥关键代偿作用，通过旁路血管通路可维持残余灌注、延缓梗死核心扩大并延长治疗时间窗[9]。侧支循环的充足程度显著影响缺血核心体积与临床预后，尤其在接受血管内血栓切除术的患者中具有决定性意义[10]。相反，侧支循环不足会加速缺血进程，并增加再灌注损伤风险，包括出血转化与神经功能恢复不良[11]。由于侧支灌注直接改变局部血流动力学状态，其代偿水平可影响基于影像学的梗死核心评估结果，例如，侧支代偿不良常导致CTP高估梗死核心体积，而mCTA则能更可靠地反映侧支灌注动力学特征并具有更高的预后价值[12]。
图 1展示了侧支代偿状态、再灌注结局与梗死进展的关系。在血管闭塞阶段，良好侧支代偿可通过维持残余灌注显著减缓梗死进展，而侧支代偿不足则表现出更快的缺血演变。当成功再灌注后，梗死进展轨迹进一步受微循环再灌注效果影响：在有效再灌注条件下，良好侧支代偿可使梗死进展趋于稳定甚至逆转，而侧支代偿不足患者仍可能出现持续性组织损伤；在无复流状态下，即便血管造影显示再通成功，缺血进展仍将持续，且侧支代偿状况决定其进展速率与最终损伤程度。
图 1 侧支代偿状态与血管再通结局共同调控梗死进展动力学的概念示意图Fig. 1 Conceptual diagram of collateral status and vascular recanalization jointly regulating dynamics of infarct progressionDuring the vessel occlusion period, good collateral circulation slows the progression of infarction, whereas poor collateral circulation leads to more rapid infarct expansion. The black dashed line indicates the time point of vascular recanalization. After recanalization, tissue perfusion outcomes may diverge into effective microvascular reperfusion or no-reflow. In the case of effective reperfusion, infarct progression tends to stabilize under good collateral conditions, whereas patients with poor collateral status may still experience ongoing tissue injury. In contrast, under no-reflow, ischemic progression continues despite angiographic evidence of successful recanalization, and the collateral status further influences the progression and the final extent of tissue damage.
下载:
全尺寸图片
除临床影像学证据外，基础研究从解剖结构、遗传背景及血流动力学等层面证实了侧支循环对AIS缺血演变的决定性作用[13]。不同小鼠品系间侧支血管的数量与直径存在显著差异，直接影响缺血后残余灌注水平与最终梗死体积；同时，内皮型一氧化氮合酶（endothelial nitric oxide synthase，eNOS）、血管内皮生长因子（vascular endothelial growth factor，VEGF）等关键信号通路的遗传调控是影响侧支血管发育与代偿能力的重要机制[13]。基础研究数据为临床观察到的侧支循环异质性提供了重要生物学依据，进一步支持对侧支循环实施精细量化评估的必要性。
2 侧支代偿能力的主要决定因素
侧支代偿能力在个体间存在显著异质性，其影响因素涵盖先天血管结构、慢性血流动力学适应及多种可调控危险因素[14]。（1）先天解剖变异（如软脑膜侧支数量、前后循环交通支发育情况）直接决定可动员的侧支通路储备，是个体侧支储备的基础[15]。研究显示，软脑膜侧支血管在基线CTA上的表现与梗死体积和临床结局密切相关[16]，不同CTA侧支评分系统（如0~3级视觉模拟量表）也部分反映了这种结构性差异[17]。（2）长期动脉粥样硬化或大血管狭窄可通过慢性低灌注诱导侧支血管重塑，增强代偿侧支供血能力；而缺乏长期适应过程的患者在急性血管闭塞时更易出现侧支募集不足[18]。此外，颅内动脉粥样硬化亚型可能调节侧支功能与预后的关联强度[19]。（3）多种可调控危险因素能够削弱侧支募集能力。高血压、糖尿病及吸烟等因素可损害内皮功能与血管舒缩反应，促使侧支血管稀疏化（collateral rarefaction），降低急性闭塞时的血流代偿水平[20]。脑白质高信号（white matter hyperintensity，WMH）作为慢性小血管病的标志，也被证实可削弱软脑膜侧支的募集能力[21]。
这些解剖、生理及代谢性因素共同导致侧支循环的高度异质性，不仅影响梗死进展速度与最终梗死体积[22]，也影响组织对再灌注后血流波动的耐受性[9]。然而，现有主流视觉定性或半定量评分体系仅能提供粗粒度离散分级，难以稳定捕捉连续性代偿差异，导致评估者间一致性及跨中心可重复性受限[23]。因此，对侧支代偿能力的精准解析，将推动评估模式从主观分级向客观、标准化、定量化表征转变，也为后续影像组学、自动化特征提取及深度学习方法的发展提供了明确方法学基础。
3 侧支循环的传统影像学评估方法及其局限性
侧支循环的影像学评估成为AIS临床风险分层、患者选择和治疗反应预测的核心环节。当前临床评估体系主要基于CTA、mCTA、DSA、磁共振血管成像（magnetic resonance angiography，MRA）与CTP等，它们从解剖结构和灌注动力学2个维度提供侧支代偿能力的重要信息，有助于判断缺血大脑半球的残余灌注能力及潜在预后。
3.1 传统影像学技术及侧支代偿评估方法
3.1.1 基于CTA的侧支循环评分
CTA是目前临床应用最广泛的非侵入性侧支评估手段，常用评分体系为Tan评分（0~3级），主要通过对比缺血侧与对侧大脑半球血管充盈程度判断侧支代偿状态[24]。DAWN试验显示，CTA评估的侧支代偿分布为Tan 3级44%、2级31%、1级22%、0级3%[25]。CTA的优势在于获取快捷、空间分辨率高并可与成像检查同步完成取栓评估[26]。
3.1.2 基于DSA的侧支代偿评估（ASITN/SIR侧支分级）
DSA仍被视为侧支循环评估的重要方法，因为其时间分辨率高，可直接观察侧支募集过程与血流方向[27]。ASITN/SIR侧支分级（1~4级）依据远端血管的充盈速度与灌注范围进行分层评估；在DAWN试验中，DSA评估的侧支代偿分布为4级7%、3级39%、2级47%、1级7%[25]。
从影像学表现来看，DSA可同时显示颅内动脉主干与皮质表面血管，为侧支代偿状态的定性判断提供了直观依据。如图 2所示，在左侧颈内动脉造影下，侧支代偿较差的病例表现为皮质血管显影稀疏、远端分支充盈受限；侧支代偿较好的病例则可见广泛、致密的皮质血管网络，覆盖更大的皮质区域，并呈逆行充盈特征。该示例直观反映了基于DSA判断侧支代偿状态高度依赖视觉经验与整体印象，缺乏可重复的定量指标。
图 2 DSA显示不同侧支代偿状态的代表性示例Fig. 2 Representative examples of collateral status assessed by DSAImages from the Neurovascular Center of The First Affiliated Hospital of Naval Medical University. Under left internal carotid artery (LICA) angiography, sparse cortical vessel visualization and limited distal branch filling suggest poor leptomeningeal collateral circulation (A, B). In contrast, a dense cortical vascular network with broad coverage and evidence of retrograde filling suggests good collateral status (C, D). The lower region of each image corresponds to the internal carotid artery trunk, while the upper region represents the cortical vascular territory, illustrating that DSA-based assessment primarily relies on the density and extent of cortical vessel opacification for qualitative evaluation of collateral status. DSA: Digital subtraction angiography.下载:
全尺寸图片
3.1.3 MRA与CTP参数
MRA可用于评估较大血管层面的供血改变及侧支相关表现[28]。CTP则通过灌注参数间接反映侧支供血状态，如HIR常被用于推测侧支功能[29]。
3.2 传统影像学侧支循环评估方法的临床应用场景
尽管传统侧支评估多以定性或半定量形式呈现，CTA/mCTA及CTP衍生灌注指标仍是目前临床与研究中应用最广泛的侧支评估工具。在AIS诊疗流程中，这些影像信息通常借助标准化后处理平台进行快速分析与呈现，从血管解剖显影到组织残余血流等层面间接反映侧支循环代偿状况，已广泛用于判断灌注状态、预测梗死演变、评估再灌注获益及筛选适宜血管内治疗或静脉溶栓的患者。
在实际应用中，CTP相关参数的获取与解释往往依赖成熟的商业化或半自动化后处理工具，如RAPID、UGuard及MIStar等[30]。这类平台能够在相对统一的算法和阈值框架下自动生成缺血核心、低灌注区及灌注延迟分布图谱，并进一步提取时间至最大残余函数延迟（time to maximum of the residue function，Tmax）、HIR等关键灌注指标，为侧支代偿状态提供可操作的量化参考。
表 1总结了具有代表性的研究工作[11, 31-36]，展示了传统影像侧支评估在真实临床研究中的应用场景。这些研究表明，基于CTP的灌注图谱及其衍生参数（如Tmax、灌注延迟模式和HIR）、CTA/mCTA的侧支分级及静脉回流特征等指标，已用于患者风险分层、组织命运预测和治疗决策支持。值得注意的是，尽管上述影像学指标来源和表现形式各异，但在当前临床实践中多作为侧支代偿能力的替代性表征，而非直接、统一的量化指标，这在一定程度限制了不同研究之间结果的可比性。而且随着AIS治疗策略的精细化及治疗时间窗的不断延长，临床对侧支灌注状态的刻画需求已从早期“是否存在侧支”，逐步转向对其空间分布特征、时间动态变化及稳定性的多维度评估。
表 1 传统CTA/mCTA/CTP指标在侧支循环评估中的临床应用研究Table 1 Clinical application of conventional CTA/mCTA/CTP indicators in collateral circulation assessmentStudy Imaging indicator Clinical role Key finding Developed and validated a convolutional neural network model to analyze CTP images for predicting 3-month functional outcome after intravenous thrombolysis Direct input: CTP perfusion maps (cerebral blood flow, cerebral blood volume, and time to peak). Comparison: automated volumes (core and penumbra) from MISTAR software Improves prediction of thrombolysis response using advanced machine learning on CTP images (1) The convolutional neural network model outperform both the original and a modified MISTAR model. (2) Combining convolutional neural network-derived imaging features with baseline clinical data further improve prediction. (3) The convolutional neural network model shows potential for aiding thrombolysis patient selection[31] Combined assessment of early ischemic changes and mCTA collateral extent for predicting EVT outcomes Non-contrast CT early ischemic score; mCTA collateral extent Evaluates ischemic severity and collateral circulation when CTP is unavailable Combined imaging scores are significantly associated with 90-d functional independence and show predictive performance comparable to CTP-derived ischemic core volume[32] Baseline CTA venous outflow characteristics for predicting EVT outcomes Single-phase CTA cortical vein opacification score Complementary indicator to arterial collateral assessment for predicting recanalization quality and functional outcomes Favorable venous outflow is associated with higher-quality recanalization and better 90-d functional outcomes, independent of arterial collateral grading[33] Evaluation of intravenous thrombolysis efficacy basedon imaging-reflected ischemictissue progression rate CTP mismatch and quantitative net water uptake Tissue-level indicator reflecting collateral compensation to optimize patient selection for thrombolysis Imaging-derived tissue progression status predicts neurological improvement and 90-d functional independence better than time window alone; patients with slower progression derive greater benefit from thrombolysis[34] Regulation of reperfusion quality and futile recanalization by leptomeningeal collateral status Leptomeningeal collateral extent and filling assessed by CTA/mCTA Identifies whether effective tissue perfusion can be achieved after recanalization and predicts the risk of futile recanalization Poor leptomeningeal collaterals increase the likelihood of futile recanalization despite successful vascular recanalization and are associated with unfavorable outcomes, whereas good collaterals significantly improve reperfusion quality and reduce futile recanalization rates[11] Comparison of imaging selection strategies for thrombectomy outcomes in late-window (6-24 h) ischemic stroke CTA collateral assessment (single-phase or multiphase CTA) +/－ CTP Evaluates whether CTA-based collateral imaging alone can replace CTP for patient selection in late-window EVT Multicenter pooled analyses show that patients selected using CTA collateral imaging achieve similar 90-d functional outcomes to those selected using CTP-based strategies, suggesting that CTA collateral assessment may serve as an alternative imaging approach for late-window EVT[35] Evaluation of the accuracy and time dependence of CTP-derived ischemic core volume estimation CTP ischemic core volume (relative cerebral blood flow, rCBF＜30% or＜20%) compared with follow-up MRI-DWI infarct volume Evaluates the reliability of CTP-derived ischemic core estimation for patient selection and prognostic prediction across different onset time windows CTP-derived ischemic core volume correlates well with follow-up infarct volume. Overestimation mainly occurs within 90 min of onset in patients with rapid reperfusion; using a stricter rCBF＜20% threshold reduces early core overestimation[36] CTA: Computed tomography angiography; mCTA: Multiphase computed tomography angiography; CTP: Computed tomography perfusion; CT: Computed tomography; EVT: Endovascular thrombectomy; rCBF: Relative cerebral blood flow; MRI-DWI: Magnetic resonance imaging diffusion-weighted imaging. 3.3 传统侧支循环评估方法的局限性
大量临床研究已充分证明侧支循环在治疗反应与预后判断中的重要作用，且部分代表性AIS研究已将侧支评估明确纳入研究设计或分析框架，用于患者入组筛选、风险分层及治疗效应修饰探索。然而，在真实世界临床实践中，侧支循环评估主要依赖于CTA、DSA或CTP等传统影像手段，虽然在临床可及性和可解释性方面具有优势，但仍存在以下不足。（1）主观评分差异大：如ASITN/SIR侧支分级和Tan评分均依赖操作者对血管充盈程度的视觉判断，评估者间一致性有限，尤其在“中等侧支”“延迟充盈”等模糊区域更易出现评分偏差[37]。（2）成像协议与后处理标准不统一：不同医院、扫描设备、采集参数（相位数、时间窗、造影剂流速等）均可影响CTA/CTP对侧支的显影效果，导致跨中心可重复性差[38]。例如，CTP的灌注参数计算受去噪算法影响显著[39]。（3）血管重建复杂且耗时：三维血管重建、中心线提取与延迟相分析流程复杂，对技术人员培训要求高，不利于在急诊取栓中快速应用[40]。研究表明，传统评估方法平均耗时15~30 min[41]。（4）缺乏全自动化、客观化指标：多数评分体系仍停留在定性与半定量层面，难以捕捉侧支灌注的空间-时间动态信息[42]，同时也限制了与预后模型或治疗策略的深度结合[43]。例如，目前尚无标准化定量参数用于预测侧支与90 d功能预后的关系[44]。综上所述，传统影像学方法对侧支的量化评估能力有限、跨中心一致性差、依赖人工判读且流程耗时较长，限制了其在精准化卒中管理中的进一步应用，也为后续探索自动化、标准化及深度学习驱动的侧支评估方法提供了明确的方法学需求。
4 深度学习在侧支循环评估中的应用
当前深度学习在侧支代偿量化评估领域的研究主要形成3条技术路径：基于影像表现的自动侧支分级、基于血管几何的结构化量化，以及融合影像与临床信息的多模态预测模型。这3类方法分别强调侧支状态的识别、侧支相关血管结构的刻画及侧支对组织命运和功能结局的影响。
4.1 深度学习侧支循环评估的核心任务与技术路径
如前所述，侧支代偿在解剖结构、血流动力学及时间演变层面均表现出显著的连续性与个体间异质性，这类复杂特征难以通过传统离散评分体系稳定刻画。深度学习方法为从影像中隐式建模这些生理与结构差异提供了新的技术路径。
4.1.1 自动侧支分级
深度学习在侧支代偿评估中最核心的应用之一是自动侧支分级，其目标是直接从原始影像中识别良好或不良侧支状态，实现Tan评分等人工视觉评分体系的自动化与客观化。
在CTA环境下，有研究采用三维卷积神经网络（convolutional neural network，CNN）对全脑三维CTA进行表征学习，通过捕捉血管显影强度、分支密度、局部纹理及灌注模式等特征，实现对侧支状态的自动判别[45]。该方法不依赖手工标注，直接从CTA体数据中提取与侧支灌注相关的空间特征，在已报道数据集中表现出较高一致性，分类性能接近人工视觉评分。其特征整合主要依赖卷积感受野逐层扩展完成，而未引入显式全局依赖建模机制，因此在结构清晰性与训练稳定性方面具有优势，但对跨区域长程依赖关系的建模能力相对依赖网络深度与参数规模。
在此基础上，有研究将侧支自动分级从单时相CTA扩展至mCTA，以利用灌注动力学中的时间维度信息。与单一时间点的静态结构相比，mCTA能够更全面地呈现侧支灌注延迟、动态填充模式及静脉期血液回流等关键生理特征，而这些特征在传统人工评分体系中往往难以稳定捕捉。基于此，mCTA深度学习模型通过融合动脉期、动静脉期、静脉期与延迟期CTA最大强度投影影像，并结合多分支特征提取与注意力机制，对跨时相灌注信息进行联合建模。与纯卷积网络相比，融入注意力机制有助于在不同时相之间进行特征重加权，增强动态信息整合能力，但整体框架仍以卷积主干为核心。在已报道的真实临床数据中，该模型整体分类准确率约为90%，优于单时相CTA模型[46]。
总体而言，在自动侧支分级任务中，卷积网络因其结构清晰、训练稳定而成为主流技术路径；注意力机制主要用于增强时间信息融合能力，而非彻底改变模型结构范式。目前尚缺乏在统一数据集上对不同架构进行系统比较的研究，各类模型的相对优势仍有待进一步验证。
4.1.2 血管结构提取与量化
在自动侧支分级基础上，更精细的侧支循环评估依赖于对血管结构的准确重建与量化，该环节是实现侧支结构-功能关联建模的重要基础。侧支灌注有效性不仅由灌注动力学决定，还受血管几何形态、远端连通性、分支结构复杂度及局部-全局血管环境制约。因此，深度学习驱动的三维血管结构提取与量化方法正逐渐成为侧支代偿评估中不可或缺的组成部分。
Li等[47]提出一种基于3D CTA的多尺度血管结构建模框架。该方法首先以不同尺度裁剪包含目标血管及邻域的三维立方体，使用基于广度优先搜索的母血管路径提取策略从CTA中保留“主血管-病灶-邻域血管”的整体结构线索；随后采用多尺度三维CNN在各空间尺度上学习血管管腔形态、走向变化、局部分支模式及局部纹理等细粒度结构特征，捕捉人工测量难以量化的几何差异；进一步引入Transformer模块，通过自注意力机制建模局部血管片段与周围结构的空间依赖性，增强对血管拓扑结构长程关联关系的表达能力。相比纯卷积结构，引入Transformer模块理论更利于捕捉复杂血管网络中的长程依赖，但其计算复杂度较高，对数据规模与计算资源的要求也更高。消融实验显示，邻域血管环境对模型的预测性能至关重要，凸显了空间上下文信息在功能预测中的重要生理学意义。
为提升复杂血管结构在不同影像模态中的一致性，有研究者提出跨模态脑血管分割框架[48]。该方法针对CTA、MRA、DSA之间显著的影像差异，通过Frangi血管增强、配准后伪标签构建及三维U-Net网络助力的跨模态学习，在无大量人工标注的情况下实现了结构一致、连通性可靠的三维血管特征提取。三维U-Net属于典型卷积结构，在局部形态建模与训练稳定性方面具有优势，但未引入显式全局依赖建模机制。研究表明，该方法在Dice相似系数及连通性感知Dice指标等方面优于主流生成对抗网络或传统伪标签技术，可较好地重建远端小血管与复杂树状结构。
上述方法分别从多尺度血管结构细化与跨模态结构一致化2个角度提升了血管几何表征质量，前者利用以多尺度三维CNN为主干并引入Transformer编码器的混合框架，强化了对长程依赖与拓扑关联的建模能力；后者依托稳定的卷积框架提升了跨模态影像的一致性。总体而言，卷积网络在结构提取任务中仍占主导地位，而引入Transformer的混合架构更多用于增强全局依赖建模能力，目前尚缺乏统一评价体系下对各类架构的直接对比分析。
4.1.3 多模态融合与预后预测
深度学习在侧支代偿研究中的第3条技术路径是构建多模态预测模型，通过联合临床特征与影像参数更全面地刻画灌注动力学、侧支状态及患者背景之间的耦合关系，从而提高对最终梗死体积和功能结局（90 d改良Rankin量表评分）的预测能力。相较于仅依赖影像或临床信息的单模态模型，多模态融合的深度学习可在高维特征空间中捕捉判别性模式，更贴近真实临床决策场景下的风险分层与预后预测需求。
基于CTP的多模态框架已展现出显著潜力。一项研究以脑血流量、脑血容量、平均通过时间、达峰时间等CTP参数图作为三维输入，利用改进型残差神经网络从CTP中学习三维时空特征，同时将年龄、美国国立卫生研究院卒中量表评分、是否接受静脉溶栓和发病至入院时间等临床变量编码为结构化特征向量，构建端到端的多模态预测系统[49]。在此基础上，该研究进一步比较了后融合、混合融合与基于动态仿射特征映射变换（dynamic affine feature map transform，DAFT）的动态跨模态特征交互机制。DAFT模块通过结合临床变量对高维影像特征进行动态缩放与偏置校正，使影像与临床语义空间实现显式对齐，提高了模型对灌注延迟、半暗带可挽救性等与侧支供血效率密切相关特征的灵敏度。在真实世界98例AIS患者数据中，采用DAFT多模态模型的最佳预测性能（AUC＝0.75、F1分数＝0.80）显著优于单一模态模型（影像参数AUC＝0.70，临床指标AUC＝0.71）及传统特征拼接策略，提示跨模态显式交互有助于释放影像与临床信息的互补价值。
从跨任务整体视角观察，多模态建模凸显了当前侧支代偿深度学习研究在模型架构上的共性。现有研究整体仍以三维CNN及其变体（如三维U-Net）为主体框架，优势在于能够稳定提取血管纹理、管腔形态及局部灌注相关空间特征，训练过程相对稳定，对中小规模医学影像数据适应性较好，因此在自动侧支分级与血管结构提取等任务中占据主导地位；局限在于对跨区域长程依赖关系的建模主要依赖感受野扩展，缺乏显式全局依赖机制。
在卷积主干基础上引入注意力机制的模型可对不同时相或空间区域进行特征重加权，尤其适用于mCTA的多时相信息融合，但整体表征仍以局部卷积特征为核心。相比之下，结合Transformer模块的混合架构通过自注意力机制显式建模全局依赖关系，在复杂血管拓扑及跨区域灌注关联刻画中具有潜在优势，但其计算复杂度较高，对数据规模与标注质量更为敏感，目前多见于样本规模相对有限的探索性研究场景。
由此可见，不同架构在建模能力、数据需求与计算开销之间呈现出显著的权衡关系：卷积网络更具稳定性与可落地性，注意力模块有助于强化多时相信息融合，而Transformer模块更有利于捕捉复杂拓扑与长程依赖关系。由于缺乏统一数据集与标准化评价体系下的直接对照研究，各类模型在侧支代偿量化评估中的相对优势仍有待进一步验证。
4.2 深度学习在侧支循环评估中的关键瓶颈与挑战
4.2.1 数据稀缺性与类别不平衡
侧支代偿相关研究中，高质量CTA、mCTA、NCCT数据有限，且“不良侧支”样本占比通常极低，导致模型在训练中难以学习少数类特征，从而降低分类性能[50]。现有研究多通过类别重加权、焦点损失函数、数据增强和半监督策略缓解这一不足，但整体数据分布不均仍是制约模型性能的主要瓶颈[51]。
4.2.2 影像特征低对比度与延迟充盈不一致
侧支血管纤细、对比度低，且延迟充盈在CTA中表现差异明显，使深度学习难以稳定捕捉其纹理与空间特征[52]。噪声、伪影及不同患者的灌注时间窗差异进一步导致侧支可视化质量不一致，影响自动分级任务的鲁棒性[53]。
4.2.3 跨中心泛化能力弱、域偏移严重
不同医院在扫描协议、造影剂注射方案、设备厂商及图像后处理算法上的差异会导致显著域偏移，使模型在外部数据上的表现明显下降[54]。尽管已有研究尝试通过多中心训练、数据标准化或自监督预训练改善泛化性，但跨平台稳定性仍不足以支持临床大规模应用[55]。
4.2.4 缺乏统一侧支分级标准
CTA、mCTA、DSA目前存在多种侧支评分体系（如Tan评分、Maas评分、ASITN/SIR侧支分级），对“良好”和“不良”侧支状态的定义差异明显，影响训练标签质量和模型评估统一性。不同中心在预处理和评分习惯上的差异也会降低模型输出可比性，成为深度学习侧支量化评估的重要限制因素[56]。
4.3 深度学习侧支循环评估模型的性能与验证现状
尽管深度学习在侧支代偿量化评估中的应用仍处于发展初期，但已有多项研究验证了其在不同影像模态和任务场景中的性能表现。
在基于mCTA的自动侧支分级任务中，有研究将4期CTA最大强度投影图像作为联合输入，通过CNN同时建模动脉期、动静脉期、静脉期及延迟期的充盈特征，实现“良好”与“不良”侧支状态的自动二分类[57]。多时相灌注动力学信息显著增强了模型对延迟充盈和慢充盈侧支的识别能力，相较于单时相CTA表现出更高的判别稳定性。
另一项基于mCTA的研究从基底节与上节水平提取代表性层面，将原始图像与最大密度投影图像组合后输入五层卷积网络，并实现“良好/中等”与“差”侧支状态的自动判别[58]。在验证集中，该模型达到0.746的准确度和0.70的AUC值，提示传统mCTA具备一定深度学习应用潜力。然而，由于样本规模有限且缺乏与人工Tan评分的一致性评估，其临床泛化能力仍需进一步验证。
除结构影像外，深度学习也被用于推断侧支的血流动力学功能。有研究利用DSC-MRI灌注图像提取放射组学特征，构建多阶段深度学习网络以预测DSA侧支三分级血供状态（差、中等、良好）[59]。模型在区分两极类别（差、良好）时表现相对可靠，但对“中等”的识别仍面临挑战，提示灌注动力学深度学习模型可部分替代DSA，尚无法完成功能性三分类等复杂度较高的任务。
在无造影剂场景下，三维Siamese网络亦被用于从NCCT中学习缺血大脑半球与对侧半球的纹理、密度及结构差异，间接推断侧支代偿状况[50]。该网络在良好/中等与不良侧支代偿的二分类任务中表现优秀，并在不同时间窗均表现稳定，表明NCCT可作为急诊条件下侧支代偿状态的间接影像学替代指标，为无造影剂环境中的自动评估提供了可行途径。
总体来看，这些研究分别基于mCTA、磁共振灌注成像和NCCT三类主要影像手段，覆盖了结构层面、灌注动力学层面及造影剂条件下侧支代偿状态的间接评估等不同任务方向。尽管研究间的任务设定、样本量与标签定义有所差异，无法直接比较具体性能，但已有证据一致支持深度学习能够自动提取与侧支灌注相关的判别性特征，量化能力已接近人工视觉评分。
5 总结和展望
深度学习为侧支代偿的客观化与自动化量化评估提供了新的方法学基础，但现有研究仍多停留在单中心验证或特定任务场景下的性能展示阶段。要推动该类方法从“研究可行”走向“临床可用”，未来研究需围绕数据标准化、灌注动力学建模、跨中心泛化能力、模型可解释与可审计输出以及与临床工作流程的系统整合等关键环节。
第一，多中心数据体系与标准化建设仍是提升可推广性的前提。当前研究普遍受数据规模有限、类别不平衡及标签定义不一致制约；同时，不同中心在扫描协议、相位设置、造影剂注射方案与后处理流程上的差异会造成显著域偏移，使模型在外部数据中表现不稳定。未来应推动多中心协作，构建覆盖不同设备厂商及真实世界临床实践场景的代表性数据队列，并在数据采集与预处理方面形成可执行的最小一致化规范；在标签层面，应尽可能明确侧支分级标准及“良好”和“不良”定义边界，减少因评分体系差异引入的系统性噪声，为跨研究比较与模型复现奠定基础。
第二，侧支循环评估的建模目标需要从静态外观分级进一步转向对灌注动力学过程的显式刻画。侧支循环的临床意义不仅体现在血管显影范围，更与延迟充盈、逆行灌注、静脉期回流等动态过程密切相关。未来研究应更充分利用多时相血管成像与灌注成像所提供的时间维度信息，将时间维度灌注动力学信息作为与结构特征同等重要的建模维度；在模型设计上，可探索面向序列的时空建模框架，将血管连通性、远端分支恢复能力及灌注延迟模式等更具生理指向性的特征纳入统一表征，以提升模型对真实侧支补偿过程的灵敏度与解释力度。
第三，跨中心泛化能力需通过更严格的外部验证与针对性域鲁棒策略支撑。仅在内部验证集中获得较高指标并不足以证明临床可用性，未来研究应优先开展跨中心、跨设备、跨协议的外部测试，系统报告不同域条件下的性能变化，并分析模型性能衰减模式及失效情境。在此基础上，可结合自监督预训练、域适配与鲁棒训练等策略降低域偏移影响，同时强化对少数类（如不良侧支状态）识别稳定性的评估，避免模型在关键高风险人群中出现性能下降。
第四，可解释性与可审计性直接决定深度学习侧支评估能否被临床接受。现有模型多以端到端分类为主，缺乏与血管解剖结构或灌注生理过程相对应的证据表达，限制了医生对模型可靠性的判断。未来可将模型输出由单一类别结果拓展为“分级结果+证据呈现”的组合形式，例如提供关键区域贡献、血管连通性变化或延迟充盈相关响应特征，并与临床可理解的影像学线索形成对应。此外，模型应具备质量控制与不确定性提示能力，在图像质量不足、域偏移显著或预测不稳定时主动提示“需人工复核”，并将相关病例纳入人机协同监督框架（human-in-the-loop supervisory framework），即由经过培训的影像医师对模型预测及其证据图进行复核与必要修正，同时将修正结果回流至模型质量监控与持续更新流程，在降低误用风险的同时提升模型的可审计性与临床可信度。
第五，深度学习侧支循环评估的价值应从“影像量化工具”进一步延伸至“临床决策支持模块”。结合现有临床研究中侧支循环评估在患者筛选、风险分层与治疗效应修饰分析中的实际作用，未来更具临床意义的方向是将侧支代偿量化结果与临床变量、再通质量及组织命运指标联合建模，用于解释梗死进展异质性、识别无效再通风险、评估治疗获益窗口及辅助个体化管理策略。同时，推理效率与流程适配同样关键，模型需在时间敏感的急诊决策场景中实现快速且稳定的推理输出，方可真正融入卒中再灌注决策链条。
未来研究重点不应局限于提升单一分类指标，而应面向真实临床异质性，构建兼具标准化数据基础、灌注动力学表征能力、跨中心稳健性、可解释输出与临床流程可嵌入性的侧支循环评估体系，推动深度学习侧支代偿评估由方法学验证阶段迈向可复制、可推广且具备临床整合能力的应用阶段。
-
图 1 侧支代偿状态与血管再通结局共同调控梗死进展动力学的概念示意图
Fig. 1 Conceptual diagram of collateral status and vascular recanalization jointly regulating dynamics of infarct progression
During the vessel occlusion period, good collateral circulation slows the progression of infarction, whereas poor collateral circulation leads to more rapid infarct expansion. The black dashed line indicates the time point of vascular recanalization. After recanalization, tissue perfusion outcomes may diverge into effective microvascular reperfusion or no-reflow. In the case of effective reperfusion, infarct progression tends to stabilize under good collateral conditions, whereas patients with poor collateral status may still experience ongoing tissue injury. In contrast, under no-reflow, ischemic progression continues despite angiographic evidence of successful recanalization, and the collateral status further influences the progression and the final extent of tissue damage.
下载:
全尺寸图片
图 2 DSA显示不同侧支代偿状态的代表性示例
Fig. 2 Representative examples of collateral status assessed by DSA
Images from the Neurovascular Center of The First Affiliated Hospital of Naval Medical University. Under left internal carotid artery (LICA) angiography, sparse cortical vessel visualization and limited distal branch filling suggest poor leptomeningeal collateral circulation (A, B). In contrast, a dense cortical vascular network with broad coverage and evidence of retrograde filling suggests good collateral status (C, D). The lower region of each image corresponds to the internal carotid artery trunk, while the upper region represents the cortical vascular territory, illustrating that DSA-based assessment primarily relies on the density and extent of cortical vessel opacification for qualitative evaluation of collateral status. DSA: Digital subtraction angiography.
下载:
全尺寸图片
表 1 传统CTA/mCTA/CTP指标在侧支循环评估中的临床应用研究
Table 1 Clinical application of conventional CTA/mCTA/CTP indicators in collateral circulation assessment
Study Imaging indicator Clinical role Key finding Developed and validated a convolutional neural network model to analyze CTP images for predicting 3-month functional outcome after intravenous thrombolysis Direct input: CTP perfusion maps (cerebral blood flow, cerebral blood volume, and time to peak). Comparison: automated volumes (core and penumbra) from MISTAR software Improves prediction of thrombolysis response using advanced machine learning on CTP images (1) The convolutional neural network model outperform both the original and a modified MISTAR model. (2) Combining convolutional neural network-derived imaging features with baseline clinical data further improve prediction. (3) The convolutional neural network model shows potential for aiding thrombolysis patient selection[31] Combined assessment of early ischemic changes and mCTA collateral extent for predicting EVT outcomes Non-contrast CT early ischemic score; mCTA collateral extent Evaluates ischemic severity and collateral circulation when CTP is unavailable Combined imaging scores are significantly associated with 90-d functional independence and show predictive performance comparable to CTP-derived ischemic core volume[32] Baseline CTA venous outflow characteristics for predicting EVT outcomes Single-phase CTA cortical vein opacification score Complementary indicator to arterial collateral assessment for predicting recanalization quality and functional outcomes Favorable venous outflow is associated with higher-quality recanalization and better 90-d functional outcomes, independent of arterial collateral grading[33] Evaluation of intravenous thrombolysis efficacy basedon imaging-reflected ischemictissue progression rate CTP mismatch and quantitative net water uptake Tissue-level indicator reflecting collateral compensation to optimize patient selection for thrombolysis Imaging-derived tissue progression status predicts neurological improvement and 90-d functional independence better than time window alone; patients with slower progression derive greater benefit from thrombolysis[34] Regulation of reperfusion quality and futile recanalization by leptomeningeal collateral status Leptomeningeal collateral extent and filling assessed by CTA/mCTA Identifies whether effective tissue perfusion can be achieved after recanalization and predicts the risk of futile recanalization Poor leptomeningeal collaterals increase the likelihood of futile recanalization despite successful vascular recanalization and are associated with unfavorable outcomes, whereas good collaterals significantly improve reperfusion quality and reduce futile recanalization rates[11] Comparison of imaging selection strategies for thrombectomy outcomes in late-window (6-24 h) ischemic stroke CTA collateral assessment (single-phase or multiphase CTA) +/－ CTP Evaluates whether CTA-based collateral imaging alone can replace CTP for patient selection in late-window EVT Multicenter pooled analyses show that patients selected using CTA collateral imaging achieve similar 90-d functional outcomes to those selected using CTP-based strategies, suggesting that CTA collateral assessment may serve as an alternative imaging approach for late-window EVT[35] Evaluation of the accuracy and time dependence of CTP-derived ischemic core volume estimation CTP ischemic core volume (relative cerebral blood flow, rCBF＜30% or＜20%) compared with follow-up MRI-DWI infarct volume Evaluates the reliability of CTP-derived ischemic core estimation for patient selection and prognostic prediction across different onset time windows CTP-derived ischemic core volume correlates well with follow-up infarct volume. Overestimation mainly occurs within 90 min of onset in patients with rapid reperfusion; using a stricter rCBF＜20% threshold reduces early core overestimation[36] CTA: Computed tomography angiography; mCTA: Multiphase computed tomography angiography; CTP: Computed tomography perfusion; CT: Computed tomography; EVT: Endovascular thrombectomy; rCBF: Relative cerebral blood flow; MRI-DWI: Magnetic resonance imaging diffusion-weighted imaging. -
[1] LITMAN M, AZARPEYKAN S, HOOD R J, et al. Shear stress targeted delivery of nitroglycerin to brain collaterals improves ischaemic stroke outcome[J]. Adv Sci (Weinh), 2025, 12(40):e06276. DOI: 10.1002/advs.202506276. [2] BROOCKS G, FAIZY T D, MEYER L, et al. Posterior circulation collateral flow modifies the effect of thrombectomy on outcome in acute basilar artery occlusion[J]. Int J Stroke, 2022, 17(7): 761-769. DOI: 10.1177/17474930211052262. [3] GENSICKE H, AL-AJLAN F, FLADT J, et al. Comparison of three scores of collateral status for their association with clinical outcome: the HERMES collaboration[J]. Stroke, 2022, 53(12): 3548-3556. DOI: 10.1161/STROKEAHA.122.039717. [4] FAIZY T D, MLYNASH M, MARKS M P, et al. Intravenous tPA (tissue-type plasminogen activator) correlates with favorable venous outflow profiles in acute ischemic stroke[J]. Stroke, 2022, 53(10): 3145-3152. DOI: 10.1161/STROKEAHA.122.038560. [5] FAIZY T D, MLYNASH M, KABIRI R, et al. The cerebral collateral cascade: comprehensive blood flow in ischemic stroke[J]. Neurology, 2022, 98(23): e2296-e2306. DOI: 10.1212/WNL.0000000000200340. [6] AKTAR M, TAMPIERI D, XIAO Y, et al. CASCADE-FSL: few-shot learning for collateral evaluation in ischemic stroke[J]. Comput Med Imaging Graph, 2025, 123: 102550. DOI: 10.1016/j.compmedimag.2025.102550. [7] LIU Y, YU Y, OUYANG J, et al. Prediction of ischemic stroke functional outcomes from acute-phase noncontrast CT and clinical information[J]. Radiology, 2024, 313(1): e240137. DOI: 10.1148/radiol.240137. [8] CAO R, LU Y, LI W, et al. Dynamic CTA-based whole-brain arterial-venous collateral assessment for predicting futile recanalization in acute ischemic stroke[J]. Aging Dis, 2025. DOI: 10.14336/AD.2025.0540. [9] MAGUIDA G, SHUAIB A. Collateral circulation in ischemic stroke: an updated review[J]. J Stroke, 2023, 25(2): 179-198. DOI: 10.5853/jos.2022.02936. [10] LENG X, NIE X, YAN H, et al. Collaterals and outcomes after endovascular treatment in acute large vessel occlusion: disparity by stroke etiologies[J]. Int J Stroke, 2025, 20(7): 864-873. DOI: 10.1177/17474930251324463. [11] BINDER N F, EL AMKI M, GLÜCK C, et al. Leptomeningeal collaterals regulate reperfusion in ischemic stroke and rescue the brain from futile recanalization[J]. Neuron, 2024, 112(9): 1456-1472.e6. DOI: 10.1016/j.neuron.2024.01.031. [12] BUSTO G, MOROTTI A, CARLESI E, et al. Pivotal role of multiphase computed tomography angiography for collateral assessment in patients with acute ischemic stroke[J]. Radiol Med, 2023, 128(8): 944-959. DOI: 10.1007/s11547-023-01668-9. [13] FODDIS M, BLUMENAU S, MUELLER S, et al. Ide copy number variant does not influence stroke severity in 2 C57BL/6J mouse models nor in humans: an exploratory study[J]. Stroke, 2025, 56(3): 725-736. DOI: 10.1161/STROKEAHA.124.049575. [14] MARUANI A, OBADIA M, FONTAINE L, et al. Hypoperfusion intensity ratio to differentiate between stroke etiologies in patients with a large vessel occlusion[J]. Eur Stroke J, 2025, 10(3): 796-803. DOI: 10.1177/23969873241306264. [15] BERGHOUT B P, SOYUPAK R F, IKRAM M K, et al. Variations in intracranial arterial anatomy of the circle of Willis and their association with arteriosclerosis in patients with ischemic cerebrovascular disease[J]. Int J Stroke, 2025, 20(7): 843-851. DOI: 10.1177/17474930251322678. [16] REESON P L, BROWN C E. Collaterals and stroke reperfusion: too few leads to too much[J]. Neuron, 2024, 112(9): 1378-1380. DOI: 10.1016/j.neuron.2024.03.024. [17] UNIKEN VENEMA S M, WOLFF L, VAN DEN BERG S A, et al. Time since stroke onset, quantitative collateral score, and functional outcome after endovascular treatment for acute ischemic stroke[J]. Neurology, 2022, 99(15): e1609-e1618. DOI: 10.1212/WNL.0000000000200968. [18] PARAMASIVAM P, CHOI S W, PODDAR R, et al. Impairment of neuronal tyrosine phosphatase STEP worsens post-ischemic inflammation and brain injury under hypertensive condition[J]. J Neuroinflammation, 2024, 21(1): 271. DOI: 10.1186/s12974-024-03227-z. [19] LUIJTEN S P R, VAN DER DONK S C, COMPAGNE K C J, et al. Intracranial carotid artery calcification subtype and collaterals in patients undergoing endovascular thrombectomy[J]. Atherosclerosis, 2021, 337: 1-6. DOI: 10.1016/j.atherosclerosis.2021.10.005. [20] LI H, FÖRSTERMANN U, XIA N, et al. Pharmacological targeting of endothelial nitric oxide synthase dysfunction and nitric oxide replacement therapy[J]. Free Radic Biol Med, 2025, 237: 455-472. DOI: 10.1016/j.freeradbiomed.2025.06.009. [21] VOORTER P H M, STRINGER M S, VAN DINTHER M, et al. Heterogeneity and penumbra of white matter hyperintensities in small vessel diseases determined by quantitative MRI[J]. Stroke, 2025, 56(1): 128-137. DOI: 10.1161/STROKEAHA.124.047910. [22] PENSATO U, DEMCHUK A M, MENON B K, et al. Cerebral infarct growth: pathophysiology, pragmatic assessment, and clinical implications[J]. Stroke, 2025, 56(1): 219-229. DOI: 10.1161/STROKEAHA.124.049013. [23] AJIBOYE N, YOO A J. Biomarkers of technical success after embolectomy for acute stroke[J]. Neurology, 2021, 97(20 Suppl 2): S91-S104. DOI: 10.1212/WNL.0000000000012800. [24] SALIM H A, LAKHANI D A, MEI J, et al. Composite arterial and venous collateral score on single-phase CTA predicts 90-day outcomes in anterior circulation large-vessel occlusion stroke[J]. J Neurointerv Surg, 2025: jnis-2025-023904. DOI: 10.1136/jnis-2025-023904. [25] LIEBESKIND D S, SABER H, XIANG B, et al. Collateral circulation in thrombectomy for stroke after 6 to 24 hours in the DAWN trial[J]. Stroke, 2022, 53(3): 742-748. DOI: 10.1161/STROKEAHA.121.034471. [26] SUMMERLIN D, WILLIS J, BOGGS R, et al. Radiation dose reduction opportunities in vascular imaging[J]. Tomography, 2022, 8(5): 2618-2638. DOI: 10.3390/tomography8050219. [27] SHABAN S, HUASEN B, HARIDAS A, et al. Digital subtraction angiography in cerebrovascular disease: current practice and perspectives on diagnosis, acute treatment and prognosis[J]. Acta Neurol Belg, 2022, 122(3): 763-780. DOI: 10.1007/s13760-021-01805-z. [28] CHA E, CHUNG H, KIM E Y, et al. Unpaired training of deep learning tMRA for flexible spatio-temporal resolution[J]. IEEE Trans Med Imaging, 2021, 40(1): 166-179. DOI: 10.1109/TMI.2020.3023620. [29] VAN HORN N, BROOCKS G, KABIRI R, et al. Cerebral hypoperfusion intensity ratio is linked to progressive early edema formation[J]. J Clin Med, 2022, 11(9): 2373. DOI: 10.3390/jcm11092373. [30] YI T, PAN Z N, LIN D L, et al. Diagnostic performance of CT perfusion in detecting contralateral aplasia of the A1 segment in acute internal carotid artery occlusion[J]. J Neurointerv Surg, 2025: jnis-2025-023461. DOI: 10.1136/jnis-2025-023461. [31] CHEN Y, TOZER D J, LIU W, et al. Prediction of response to thrombolysis in acute stroke using neural network analysis of CT perfusion imaging[J]. Eur Stroke J, 2023, 8(3): 629-637. DOI: 10.1177/23969873231183206. [32] TANAKA K, KAVEETA C, PENSATO U, et al. Combining early ischemic change and collateral extent for functional outcomes after endovascular therapy: an analysis from AcT trial[J]. Stroke, 2024, 55(7): 1758-1766. DOI: 10.1161/STROKEAHA.123.046056. [33] FAIZY T D, KABIRI R, CHRISTENSEN S, et al. Association of venous outflow profiles and successful vessel reperfusion after thrombectomy[J]. Neurology, 2021, 96(24): e2903-e2911. DOI: 10.1212/WNL.0000000000012106. [34] BROOCKS G, MEYER L, BECHSTEIN M, et al. Investigating neurologic improvement after Ⅳ thrombolysis: the effect of time from stroke onset vs imaging-based tissue clock[J]. Neurology, 2023, 101(17): e1678-e1686. DOI: 10.1212/WNL.0000000000207714. [35] ALMEKHLAFI M A, THORNTON J, CASETTA I, et al. Stroke imaging prior to thrombectomy in the late window: results from a pooled multicentre analysis[J]. J Neurol Neurosurg Psychiatry, 2022, 93(5): 468-474. DOI: 10.1136/jnnp-2021-327959. [36] SARRAJ A, CAMPBELL B C V, CHRISTENSEN S, et al. Accuracy of CT perfusion-based core estimation of follow-up infarction: effects of time since last known well[J]. Neurology, 2022, 98(21): e2084-e2096. DOI: 10.1212/WNL.0000000000200269. [37] YANG W, SOOMRO J, JANSEN I G H, et al. Collateral capacity assessment: robustness and interobserver agreement of two grading scales and agreement with quantitative scoring[J]. Clin Neuroradiol, 2023, 33(2): 353-359. DOI: 10.1007/s00062-022-01216-4. [38] KIM B, YOU S H, JUNG S C. A multicenter survey of acute stroke imaging protocols for endovascular thrombectomy[J]. Neurointervention, 2021, 16(1): 20-28. DOI: 10.5469/neuroint.2020.00199. [39] PAN K, WANG H, CHEN X, et al. Comparative analysis of two mathematical algorithms for the calculation of computed tomography perfusion parameters in the healthy and diseased pancreas[J]. J Appl Clin Med Phys, 2022, 23(2): e13488. DOI: 10.1002/acm2.13488. [40] TOMASELLO A, HERNÁNDEZ D, LI J, et al. Modeling robotic-assisted mechanical thrombectomy procedures with the CorPath GRX robot: the core-flow study[J]. AJNR Am J Neuroradiol, 2024, 45(6): 721-726. DOI: 10.3174/ajnr.A8205. [41] WEISS D, RUBBERT C, KASCHNER M, et al. Prehospital telemedicine support for urban stroke care: Analysis of current state of care and conceptualization[J]. BMC Emerg Med, 2024, 24(1): 224. DOI: 10.1186/s12873-024-01142-3. [42] GOMEZ C R. Time is brain: the stroke theory of relativity[J]. J Stroke Cerebrovasc Dis, 2018, 27(8): 2214-2227. DOI: 10.1016/j.jstrokecerebrovasdis.2018.04.001. [43] XU Y, YANG J, GAO X, et al. Quantitative assessment of collateral time on perfusion computed tomography in acute ischemic stroke patients[J]. Front Neurol, 2023, 14: 1230697. DOI: 10.3389/fneur.2023.1230697. [44] VARELA R, GONZALEZ-ORTIZ F, DIAS A, et al. Plasma brain-derived tau in prognosis of large vessel occlusion ischemic stroke[J]. Stroke, 2024, 55(9): 2353-2358. DOI: 10.1161/STROKEAHA.123.046117. [45] SU J, WOLFF L, VAN ES A C G M, et al. Automatic collateral scoring from 3D CTA images[J]. IEEE Trans Med Imaging, 2020, 39(6): 2190-2200. DOI: 10.1109/TMI.2020.2966921. [46] TAN D, LIU J, CHEN S, et al. Automatic evaluating of multi-phase cranial CTA collateral circulation based on feature fusion attention network model[J]. IEEE Trans Nanobioscience, 2023, 22(4): 789-799. DOI: 10.1109/TNB.2023.3283049. [47] LI P, LIU Y, ZHOU J, et al. A deep-learning method for the end-to-end prediction of intracranial aneurysm rupture risk[J]. Patterns (N Y), 2023, 4(4): 100709. DOI: 10.1016/j.patter.2023.100709. [48] GUO Z, FENG J, LU W, et al. Cross-modality cerebrovascular segmentation based on pseudo-label generation via paired data[J]. Comput Med Imaging Graph, 2024, 115: 102393. DOI: 10.1016/j.compmedimag.2024.102393. [49] BORSOS B, ALLAART C G, VAN HALTEREN A. Predicting stroke outcome: a case for multimodal deep learning methods with tabular and CT perfusion data[J]. Artif Intell Med, 2024, 147: 102719. DOI: 10.1016/j.artmed.2023.102719. [50] AKTAR M, XIAO Y, TEHRANI A K Z, et al. SCANED: Siamese collateral assessment network for evaluation of collaterals from ischemic damage[J]. Comput Med Imaging Graph, 2024, 113: 102346. DOI: 10.1016/j.compmedimag.2024.102346. [51] ROUGÉ P, CONZE P H, PASSAT N, et al. Guidelines for cerebrovascular segmentation: managing imperfect annotations in the context of semi-supervised learning[J]. Comput Med Imaging Graph, 2025, 119: 102474. DOI: 10.1016/j.compmedimag.2024.102474. [52] YI T, SUI Y, ZHENG D, et al. Diagnostic performance of carotid ring sign on CT-angiography in internal carotid true occlusion[J]. Stroke, 2024, 55(4): 1025-1031. DOI: 10.1161/STROKEAHA.123.045156. [53] ZHANG J, HUANG C, LOK U W, et al. Enhancing row-column array (RCA)-based 3D ultrasound vascular imaging with spatial-temporal similarity weighting[J]. IEEE Trans Med Imaging, 2025, 44(1): 297-309. DOI: 10.1109/TMI.2024.3439615. [54] AYOUB M, LIAO Z, LI L, et al. HViT: hybrid vision inspired transformer for the assessment of carotid artery plaque by addressing the cross-modality domain adaptation problem in MRI[J]. Comput Med Imaging Graph, 2023, 109: 102295. DOI: 10.1016/j.compmedimag.2023.102295. [55] CAI H, ZHANG Q, LONG Y. Prototype-guided multi-scale domain adaptation for Alzheimer's disease detection[J]. Comput Biol Med, 2023, 154: 106570. DOI: 10.1016/j.compbiomed.2023.106570. [56] GU R, WANG G, LU J, et al. CDDSA: Contrastive domain disentanglement and style augmentation for generalizable medical image segmentation[J]. Med Image Anal, 2023, 89: 102904. DOI: 10.1016/j.media.2023.102904. [57] WANG J, TAN D, LIU J, et al. Merging multiphase CTA images and training them simultaneously with a deep learning algorithm could improve the efficacy of AI models for lateral circulation assessment in ischemic stroke[J]. Diagnostics (Basel), 2022, 12(7): 1562. DOI: 10.3390/diagnostics12071562. [58] HUANG C C, CHIANG H F, HSIEH C C, et al. Using deep-learning-based artificial intelligence technique to automatically evaluate the collateral status of multiphase CTA in acute ischemic stroke[J]. Tomography, 2023, 9(2): 647-656. DOI: 10.3390/tomography9020052. [59] TETTEH G, NAVARRO F, MEIER R, et al. A deep learning approach to predict collateral flow in stroke patients using radiomic features from perfusion images[J]. Front Neurol, 2023, 14: 1039693. DOI:10.3389/fneur. 2023.1039693.
