基于改进的3D U-Net骨盆CT影像多类分割

Pelvic CT images multiclass segmentation based on improved 3D U-Net

骨盆CT影像精确分割是骨盆骨疾病的临床诊断和手术规划中非常重要的环节。针对目前2D骨盆分割方法对三维医学影像进行切片处理时损失空间信息的问题,提出了改进3D U-Net网络实现对骨盆CT影像3D自动分割。实验数据为公开数据集CTPelvic1K共1184名患者骨盆CT影像,其中包含骶骨、左髋骨、右髋骨和腰椎四个部位标签。以3D U-Net骨干网络为基础,结合自注意力机制提出3D多类分割模型3D Trans U-Net,并使用迁移学习训练3D U-Net、V-Net、Attention U-Net作为对照实验。实验结果表明:3D Trans U-Net在测试集上整个骨盆区域、骶骨、左髋骨、右髋骨、腰椎Dice系数分别达到97.99%,96.70%,97.96%,97.95%,96.89%;Dice系数、豪斯多夫距离等评价指标均优于现有经典网络3D U-Net、V-Net、Attention U-Net。因此,改进的3D Trans U-Net对骨盆不同部位具有较好的分割效果,为精准医治骨盆骨疾病提供了一条有效的技术途径。

The accurate segmentation of pelvic CT images is a very important step in the clinical diagnosis and surgical planning of pelvic bone diseases.In allusion to the loss of spatial information in the existing 2D pelvic segmentation method when slicing 3D medical images,an improved 3D U-Net network is proposed to realize 3D automatic segmentation of pelvic CT images.The experimental data is the pelvic CT images of 1184 patients in the open dataset CTPelvic1K,including four parts:labels of sacrum,left hipbone,right hipbone and lumbar vertebra.Based on the 3D U-Net backbone network,a 3D multi-class segmentation model 3D Trans U-Net is proposed by means of the self-attention mechanism,and the transfer learning is used to train3D U-Net,V-Net,and Attention U-Net as the control experiment.The experimental results show that the Dice coefficients of the whole pelvis,sacrum,left hipbone,right hipbone and lumbar vertebra of 3D Trans U-Net on the test set can reach 97.99%,96.70%,97.96%,97.95%and 96.89%,respectively.The evaluation index of Dice coefficient,Hausdorff distance,etc.are better than the existing classic networks:3D U-Net,V-Net and Attention U-Net.Therefore,3D Trans U-Net has better segmentation effect on different parts of the pelvis,providing an effective technical way for the accurate treatment of pelvic bone diseases.