基于Dense V-Network神经网络的女性肠道自动分割研究

目的 用精准自动勾画女性肠道器官的Dense V-Network模型对宫颈癌患者进行训练并评估。方法 将Dense Net与V-Net2个网络模型进行融合,形成一种用于三维CT图像自动分割的Dense V-Network算法。160例宫颈癌患者CT数据被随机分为训练集130例用于调整模型参数,测试集30例用于评估自动分割效果。采用戴斯相似性系数(DSC)等8个参数定量评估分割效果。结果 小肠DSC、杰卡德距离、体积差异性系数、敏感性指数、包容性指数、豪斯多夫距离、轮廓平均差异、质心偏差分别为0.86±0.03、0.25±0.04、0.10±0.07、0.88±0.05、0.85±0.05、(2.98±0.61) cm、(2.40±0.45) mm、(4.13±1.74) mm,结果优于单一算法(均P<0.05)。结论 Dense V-Network算法可较为准确地分割肠道器官,医生修改审查简单易行,可用于临床。

Research on automatic segmentation of female bowel based on Dense V-Network neural network

Objective To resolve the issue of poor automatic segmentation of the bowel in women with pelvic tumors, a Dense V-Network model was established, trained and evaluated to accurately and automatically delineate the bowel of female patients with pelvic tumors. Methods Dense Net and V-Net network models were combined to develop a Dense V-Network algorithm for automatic segmentation of 3D CT images. CT data were collected from 160 patients with cervical cancer, 130 of which were randomly selected as the training set to adjust the model parameters, and the remaining 30 were used as test set to evaluate the effect of automatic segmentation. Results Eight parameters including Dice similarity coefficient (DSC) were utilized to quantitatively evaluate the segmentation effect. The DSC value, JD,ΔV, SI, IncI, HD (cm), MDA (mm), and DC (mm) of the small intestine were 0.86±0.03,0.25±0.04,0.10±0.07,0.88±0.05,0.85±0.05,2.98±0.61,2.40±0.45 and 4.13±1.74, which were better than those of any other single algorithm. Conclusion Dense V-Network algorithm proposed in this paper can deliver accurate segmentation of the bowel organs. It can be applied in clinical practice after slight revision by physicians.