摘要
颅内动脉瘤是脑血管的病理性扩张,一旦破裂致死率极高.人工检测颅内动脉瘤繁琐耗时,因此有必要引入自动化检测技术.为解决现存算法在处理点云数据时空间利用率低,难以同时捕捉局部细节与整体结构的问题,设计基于片序列注意力的颅内动脉瘤点云分割算法.利用空间填充曲线将点云序列化,改善了传统点云在提取局部结构信息时的效果.提出更加精细化的片注意力,利用片分组和片交互,进一步加强模型对不同空间关系的泛化能力.同时设计网格池化操作,解决了传统算法对于信息密度和重叠不可控的问题.该方法在IntrA数据集上获得了健康血管段IoU为95.37%、动脉瘤段IoU为84.67%的分割精度,远优于大多数现存算法.
Intracranial aneurysms refer to the pathological expansion of blood vessels in the brain.Once an aneurysm ruptures,the fatality rate is extremely high.Manual detection of intracranial aneurysms is cumbersome and timeconsuming,making the introduction of automated detection technology necessary.To address the insufficient space utilization of three-dimensional(3D)point cloud data by existing algorithms and the difficulty in fully capturing both local details and overall structural information,a segmentation algorithm based on patch serialization attention for intracranial aneurysm point clouds is designed.Space-filling curves are used to serialize point clouds,which improves the effectiveness of traditional point clouds in extracting relationships between different local structures.A more refined patch attention mechanism is introduced,with patch grouping and patch interaction,to strengthen the model’s ability to generalize spatial relationships.The issues related to uncontrollable information density and overlap in traditional pooling methods are resolved through graph pooling operations.On the IntrA dataset,the segmentation accuracy achieved is 95.37%IoU for healthy blood vessels and 84.67%IoU for aneurysm segments,significantly outperforming most existing algorithms.
作者
李硕
郭莉琳
LI Shuo;GUO Li-Lin(Department of Information,School of Medical Technology and information Engineering,Zhejiang Chinese Medical University,Hangzhou 310053,China)
出处
《计算机系统应用》
2025年第7期163-173,共11页
Computer Systems & Applications
基金
浙江中医药大学校级科研项目(781100E036)。
关键词
颅内动脉瘤
医学图像处理
3D点云
语义分割
序列化
深度学习
片注意力
intracranial aneurysm
medical image processing
3D point cloud
semantic segmentation
serialization
deep learning
patch attention
