基于卷积神经网络的颅内囊状动脉瘤半自动分割模型的构建与验证研究

An semi-automatic segmentation model for intracranial saccular aneurysms based on convolutional neural networks:construction and verification

目的基于卷积神经网络创建一种半自动的颅内囊状动脉瘤分割技术。方法回顾性连续纳入2017年7月至2020年7月“中国颅内动脉瘤计划”数据库中首都医科大学宣武医院的单中心数据,所有数据在分析前均进行了匿名处理。收集所有患者的基线资料,包括性别、年龄(≥60岁和<60岁)和DSA机型、DSA序列数及动脉瘤信息,包括动脉瘤个数、直径(≥5 mm和<5 mm)、颈宽(宽颈、窄颈)及位置(分叉部、侧壁)。根据8∶1∶1的比例将数据通过随机数字表法随机分为训练集、测试集和验证集。3个数据集患者的DSA三维断层数据均采用三维旋转DSA模式在造影机完成,并由3位有经验的神经外科医师对DSA三维断层数据显示的动脉瘤进行标注,并最终生成动脉瘤的标准标签。动脉瘤分割模型包括训练阶段和分割阶段。训练阶段,使用训练集的DSA三维断层图像数据与动脉瘤的分割标签以及通过Marching Cubes算法提取的血管边缘信息,对模型进行端到端的训练,在测试集上监控模型的分割指标,保留分割指标最高的模型。分割阶段,医师在验证集的动脉瘤DSA三维断层图像上选择一个动脉瘤内部的点,截取感兴趣体积(VOI),输入训练好的血管与动脉瘤分割最优模型,得到动脉瘤的分割结果,将分割的VOI定位回原始DSA三维断层图像以获得最终的动脉瘤轮廓。将分割网络模型的分割结果与人工获取的标准标签进行比较,以计算Dice相似系数(DSC)。对验证集数据按照动脉瘤直径、颈宽、位置进行分层,以比较不同亚组间的DSC。计算动脉瘤分割掩膜的长、宽和高的边界框,将其中的最大值作为动脉瘤的最大直径,与标准标签中的最大直径进行对比。在验证集中统计并比较颅内动脉瘤标准标签人工获取时间与分割网络模型获取时间(从定位动脉瘤到获取满意的动脉瘤颈分割时间)。结果最终纳入了756例患者的969个DSA序列显示的1094个动脉瘤的三维断层数据。其中,训练集纳入604例患者共783个DSA序列的877个动脉瘤,测试集纳入77例患者共100个DSA序列的117个动脉瘤,验证集纳入75例患者共86个DSA序列的100个动脉瘤。(1)各数据集基线比较结果显示,动脉瘤直径(P=0.003)、动脉瘤位置(P=0.003)的各数据集间的差异有统计学意义。余基线资料各数据集间差异无统计学意义(均P>0.05)。(2)验证集中动脉瘤分割的平均DSC为0.868±0.078。直径≥5 mm的动脉瘤分割的平均DSC高于直径<5 mm的动脉瘤(0.891±0.041比0.855±0.088,P=0.038)。窄颈、宽颈、分叉、侧壁动脉瘤分割的DSC值分别为0.882±0.065、0.859±0.085、0.876±0.072及0.863±0.080,组间差异均无统计学意义(均P>0.05)。(3)动脉瘤分割模型在验证集所得到的掩膜最大直径与人工分割获得的标准标签的最大直径有较好的一致性[(5.78±3.18)mm比(5.37±2.92)mm,r=0.97]。在验证集中,人工分割与应用神经网络分割动脉瘤的平均时长分别为2.5 min、34 s。结论本研究基于卷积神经网络创建半自动的颅内囊状动脉瘤分割技术可较为准确分割动脉瘤,该模型有助于动脉瘤形态学分析。

Objective To create a semi-automatic technology based on convolutional neural networks for saccular aneurysm segmentation.Methods The single-center data of Xuanwu Hospital of Capital Medical University in the database of“China Intracranial Aneurysm Program”from July 2017 to July 2020 were retrospectively included,and all data were anonymized before analysis.Baseline data were collected from all patients,including sex,age(≥60 years and<60 years),DSA model,number of DSA sequences,and aneurysm information,including the number of aneurysms,diameter(≥5 mm and<5 mm),neck width(wide neck,narrow neck),and location(bifurcation,sidewall).According to the ratio of 8∶1∶1,the data were randomly divided into training set,test set and validation set by random number table method.The DSA 3D tomography data of all patients were completed in the contrast machine using the 3D rotary DSA mode,and the aneurysms shown in the DSA 3D tomography data were annotated by 3 experienced neurosurgeons,and the standard label of the aneurysm was finally generated.The proposed aneurysm segmentation method consisted of a training stage and a segmentation stage.In the training stage,the model was trained end-to-end by using the DSA 3D tomography image data of the training set,the segmentation label of the aneurysm and the vascular edge information extracted by the Marching Cubes algorithm,and the segmentation index of the model was monitored on the test set to retain the model with the highest segmentation index.In the segmentation stage,the physician selects a point inside the aneurysm on the DSA 3D tomography image of the aneurysm in the validation set,intercepts the volume of interest(VOI),inputs the trained optimal model of vascular and aneurysm segmentation,obtains the segmentation result of the aneurysm,and locates the segmented VOI back to the original DSA 3D tomography image to obtain the final aneurysm outline.The segmentation results of the segmentation network model were compared with standard labels to calculate the Dice similarity coefficient(DSC).The validation set data was stratified by aneurysm diameter,neck width,and location to compare the segmentation results in different datasets.We calculated the bounding boxes for the length,width,and height of the aneurysm segmentation mask,and used the maximum of these as the longest diameter of the aneurysm compared to the maximum diameter in the standard label.In the validation set,the standard label manual acquisition time was counted and compared with the segmentation network model acquisition time(from the time of locating the aneurysm to obtaining a satisfactory aneurysm neck segmentation).Results Finally,969 DSA sequences from 756 patients were included to show 3D tomographic data for 1094 aneurysms.Among them,604 patients with 877 aneurysms with a total of 783 DSA sequences were included in the training set,117 aneurysms with a total of 100 DSA sequences in 77 patients were included in the test set,and 100 aneurysms with a total of 86 DSA sequences were included in 75 patients in the validation set.(1)The baseline comparison results of each dataset showed that there were statistically significant differences between the datasets of aneurysm diameter(P=0.003)and aneurysm location(P=0.003).There was no significant difference between the remaining baseline data sets(all P>0.05).(2)The mean DSC of centralized aneurysm segmentation was 0.868±0.078.The mean DSC of aneurysm segmentation≥5 mm diameter was higher than that of aneurysms with<5 mm diameter(0.891±0.041 vs.0.855±0.088,P=0.038).The DSC values of narrow-necked,wide-necked,bifurcated and lateral wall aneurysms were 0.882±0.065,0.859±0.085,0.876±0.072 and 0.863±0.080,respectively,and there was no significant difference between the groups(all P>0.05).(3)The maximum diameter of the mask obtained by the aneurysm segmentation model in the validation set was in good agreement with the maximum diameter of the standard label obtained by manual segmentation([5.78±3.18]mm vs.[5.37±2.92]mm,r=0.97).In the validation set,the average time of manual segmentation and neural network segmentation of aneurysms was 2.5 min and 34 s,respectively.Conclusion In this study,a semi-automatic saccular aneurysm segmentation technique based on convolutional neural network can accurately segment aneurysms and is helpful to improve aneurysm morphology analysis.