An semi-automatic segmentation model for intracranial saccular aneurysms based on convolutional neural networks:construction and verification
10.3969/j.issn.1672-5921.2024.09.001
- VernacularTitle: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
- Author:
Jiewen GENG
1
;
Simin WANG
;
Peng HU
;
Chuan HE
;
Hongqi ZHANG
Author Information
1. 100020 首都医科大学附属北京朝阳医院神经外科
- Keywords:
Intracranial saccular aneurysm;
Segmentation model;
Neural network;
U-shape network;
Dice similarity coefficient
- From:
Chinese Journal of Cerebrovascular Diseases
2024;21(9):577-586
- CountryChina
- Language:Chinese
-
Abstract:
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 1 094 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.
