毕业生知识库

脑血管精准分割是进行脑血管疾病定量诊断和分析的重要前提，也是脑血管相关的神经外科手术导航的关键步骤，对于理解脑血管疾病的发病机制、术前诊断和术中治疗具有重要意义。本文以时间飞跃法-磁共振血管造影 (Time of Flight-Magnetic Resonance Angiography, TOF-MRA) 为研究数据，从模型驱动的统计方法以及数据驱动的半监督深度学习方法两方面对脑血管分割进行了深入研究。

当前脑血管分割方法主要可分为模型驱动和数据驱动两大类，前者主要包括统计模型为主的数学建模方法，后者则由深度学习方法组成。一方面，尽管统计模型在脑血管分割任务上取得了充分的进展，但由于来自不同站点、成像设备以及扫描参数的图像之间存在较大差异，因此难以设计通用的统计模型对数据分布进行准确的建模。此外，脑血管的极小占比和稀疏分布也对参数估计带来了困难，限制了这类方法的准确性和泛化性；另一方面，尽管深度学习在医学图像处理领域取得了显著进步，但在脑血管分割任务中始终面临着高质量标注数据不足的问题，限制了分割精度的提升，以及向不同站点、不同扫描参数和不同物种图像泛化的能力。

针对上述问题，本文从统计模型驱动和数据驱动的脑血管分割两方面展开研究，主要工作内容和创新点如下：

1. 提出了一种基于统计模型的脑血管分割方法FFCM-MRF。该方法首先利用快速模糊C均值聚类（FFCM）进行脑血管的初始分割，FFCM不仅高效准确，而且模型设计简单，参数估计过程不受脑血管稀疏分布的影响。考虑到脑血管的三维管状特性，进一步使用多尺度血管增强滤波方法，得到包含血管形状先验信息的血管增强图。然后将血管增强图嵌入到马尔可夫随机场（MRF）中，并结合空间邻域相似性约束，设计了一个新的MRF能量函数用于初始分割结果的优化。该过程不仅可以对FFCM的初始分割结果进行去噪，还可以进一步分割出更多的血管，增强血管的拓扑连通性。实验结果表明，FFCM-MRF在不同尺度和类别的数据集都展现出了较强的鲁棒性和泛化性，相比其他统计方法，FFCM-MRF具有更高的精度和更强的抗噪能力；相比深度学习方法，FFCM-MRF具有更高的效率和更强的泛化能力。
2. 提出了一种基于低质量标签优化和一致性正则化的半监督脑血管分割方法CL-MT。该方法综合有效利用少量的高质量标注数据和从FFCM-MRF分割得到的大量低质量标注数据进行高效鲁棒的学习。通过引入置信学习，CL-MT能够识别低质量标签中的错误标注体素得到误差图，依据标注误差图迭代式地优化低质量标签，从而为网络提供更准确的监督信号；一致性正则化策略则将一致性损失的计算限制在错误标注的体素区域，鼓励学生网络和教师网络在这些高不确定区域的预测趋向一致，从而学习到更鲁棒有效的特征表示。我们构建了包含人类和猕猴的多中心异质数据集，并从骨架网络、半监督策略、整体模型三方面进行了对比实验，实验结果表明该方法能够有效提高脑血管分割的精度和泛化能力，并且优于同类脑血管分割方法。

本文提出的模型驱动和数据驱动的TOF-MRA图像脑血管分割方法，实现了精准、高效的三维脑血管自动分割，有望辅助神经外科医生进行脑血管相关手术规划和脑血管疾病预防诊治。

Accurate cerebrovascular segmentation is an essential prerequisite for quantitative diagnosis and analysis of cerebrovascular diseases, as well as a crucial step in neurosurgical navigation. It holds great significance in exploring the pathogenesis of cerebrovascular diseases, providing preoperative diagnoses, and guiding intraoperative treatments. This study focuses on the model-driven statistical method and data-driven semi-supervised method for cerebrovascular segmentation, based on TOF-MRA (Time of Flight-Magnetic Resonance Angiography) data.

Cerebrovascular segmentation methods can be mainly categorized into two primary groups: model-driven and data-driven approaches. The model-driven method encompasses mathematical modeling techniques that predominantly rely on statistical models, whereas the data-driven method comprises deep learning methodologies. While statistical models have made significant progress in cerebrovascular segmentation tasks, they are often challenged by the variability of images from different sites, scanner platforms, and acquisition parameters, making it difficult to accurately model the data distribution. The extremely small proportion and sparse distribution of cerebral vessels also pose difficulties for parameter estimation, thereby limiting the accuracy and generalizability of these methods. On the other hand, deep learning has made remarkable progress in medical image processing. However, it has always been constrained by the lack of high-quality annotations for cerebrovascular segmentation, which hinders the improvement of segmentation accuracy and generalizability for images from different sites, acquisition parameters, and species.

Aiming at the problems mentioned above, this study conducted research on statistical model-driven and data-driven cerebrovascular segmentation task. The main contributions and innovations of this study are as follows:

(1) We proposed a statistical modeling-based cerebrovascular segmentation method called FFCM-MRF. The method first uses fast fuzzy c-means (FFCM) clustering for rudimentary cerebrovascular segmentation. FFCM is not only efficient and accurate, but also has a simple design and is not affected by the sparse distribution of cerebrovascular structures. Given the tubular characteristics of cerebrovascular structures, a multi-scale vessel enhancement filtering method is further applied to obtain the vessel-enhanced images that incorporate shape prior of blood vessels. Markov process was then employed to further refine the rudimentary segmentation results derived from FFCM method. This process utilizes a Markov Random Field (MRF) with a novel energy function that integrates spatial constraints among neighboring voxels. These constraints incorporate both label fields from FFCM and shape features from the vessel-enhanced images. The application of the Markov process not only reduces the noise in the label field but also enriches the vascular structure, thereby enhancing the overall segmentation performance. Experimental results demonstrated that FFCM-MRF method exhibits robustness performance across datasets of varying scales and species. It not only achieves higher accuracy and stronger noise robustness compared to statistical methods, but also outperforms deep learning methods in terms of efficiency and generalizability.

(2) We proposed a semi-supervised cerebrovascular segmentation method called confident learning-based mean teacher framework (CL-MT), which utilizes a small amount of high-quality labeled data and a large amount of low-quality labeled data obtained from FFCM-MRF. This framework builds upon the mean-teacher architecture, along with synergistic noisy label refinement and ambiguity-guided consistency regularization modules. We adapt the confident learning to refine noisy labels, thereby reducing mislabeled voxels and providing more accurate supervision for training. Furthermore, we preserved predictions on noisy regions characterized by confident learning for consistency loss, which encouraged perturbed stability in these ambiguous-yet-informative regions, driving the model to learn useful representations from unlabeled data. We conducted comprehensive experiments on heterogeneous multi-center datasets comprising images from different scanners for both humans and non-human primates. Evaluations were performed from three aspects: backbone networks, semi-supervised strategies and overall models. Compared to state-of-the art supervised and semi-supervised methods, our model achieved superior segmentation performance and generalizability.

In conclusion, the model-driven and data-driven methods proposed by this study achieved accurate and efficient three-dimensional automatic segmentation of cerebral vessels. The proposed methods will assist neurosurgeons in cerebrovascular surgical planning, diagnosing and treating of cerebrovascular diseases.