基于球B样条与空间图深度学习的三维颅内动脉瘤自动检测算法研究

颅内动脉瘤(IAs)破裂会导致严重脑中风，传统IAs检测方法囿于数据成像方式不同，样本分布不均衡且训练集合较小，使其在临床与筛查不通用。本项目创新地将IAs检测视为形状分析问题，将脑血管分割重构成三维模型后，在几何模型上实现IAs检测，打破数据成像方式与分布屏障，贯通临床与筛查应用。研究将三维模型深度学习转化为希尔伯特空间泛函优化问题，结合数字几何表示与统计机器学习方法，构建融合空间结构信息和属性信息的形质互嵌卷积核，探索基于子结构塌缩的归约策略实现图的结构优化。基于脑血管球B样条表达拟合，变形求交，生成大规模合成样本集，提高样本数量。利用智能优化方法解决几何中的拟合，求交与交线平滑等非线性非凸问题。本研究不仅可为脑血管病的诊断预防提供技术支持，也可为开发具有我国自主知识产权的脑图像处理软件提供服务。

The intracranial aneurysms rupture can cause the serious stroke, which relate to decline of daily life ability in the elderly. The existing automatic intracranial aneurysms detection method through medical images is not so effective, which is restricted to the inconsistence of data format, the imbalance of the sample distribution and small training dataset, limiting the usage of the detection method in clinical and targeted screening. This research innovatively presents that the intracranial aneurysms detection is a shape analysis problem. We detect the intracranial aneurysms in three dimension cerebrovascular mesh model after the segmentation and reconstruction of the brain vessel from the medical image. This method can break the barriers of the data format and data distribution, which can service both in clinical and screening. Then based previous research of Ball B-Spline Curves (BBSCs) representation of vessels, we study the fitting method, the deformation method and the interaction method to build the large-scale dataset of synthetic 3D aneurysm models. In theory, we present the novel spatial deep learning model, with constructing the convolution operator based on semantic and structural feature co-embedding and the pooling operator with graph collapsing and subdivision algorithm. The new learning model can be thought as the functional space in the Hilbert space. We combine the geometric idea in the statistic machine learning method and do want to give the mathematical explanation. In methodology, we use the intelligent optimization method to solve the traditional geometric problem such as fitting, interacting and cutting curve smoothing, which are the nonlinear and no-convex problem. In application, our research can not only greatly support the work of prevent, diagnose and treatment in the cerebrovascular disease, but also can service for the medical image processing platform with Chinese intellectual property rights.