基于2D/3D U-plus-net的心脏自动分割

Automatic heart segmentation based on 2D/3D U-plus-net

目的:利用2D/3D U-plus-net提高心脏自动分割的准确率。方法:收集郑州大学第一附属医院60例患者胸部扫描CT图像(数据A)及中国科学技术大学附属第一医院45例患者胸部扫描CT图像(数据B)。基于改进的AlexNet将CT图像分为两类:心脏CT图像和无心脏CT图像。在2D/3D U-net拓扑结构基础上,通过减小网络深度、在长连接中增加新节点、增加解码器中卷积次数的方法,得到改进后的2D/3D U-plus-net;将靠近腹部的心脏CT图像(图像张数由预实验决定)输入3D U-plus-net,其余图像输入2D U-plus-net;采用5倍交叉验证法对模型进行训练及测试。最后通过Dice系数、HD95和平均表面距离(MSD)评估自动分割精度。结果:数据A自动分割的Dice系数为0.941±0.012,MSD为(3.918±0.201)mm,HD95为(5.863±0.561)mm;数据B自动分割的Dice系数为0.934±0.014,MSD为(4.112±0.320)mm,HD95为(6.035±0.659)mm。结论:基于2D/3D U-plus-net的分割方法提高了心脏自动分割准确率。

Objective To improve the accuracy of automatic heart segmentation using 2D/3D U-plus-net.Methods The chest CT images of 60 patients from the First Affiliated Hospital of Zhengzhou University(Data A)and the chest CT images of 45 patients from the First Affiliated Hospital of University of Science and Technology of China(Data B)were collected.A modified AlexNet was used to divide all CT images into two types,namely heart CT images and no-heart CT images.Based on the topological structure of 2D/3D U-net,a modified 2D/3D U-plus-net was obtained by reducing network depth,increasing nodes in a long connection and increasing the convolution number of the decoder.The heart CT images near the abdomen(the number of CT images was determined by pre-experiment)were input into 3DU-plus-net,while the other heart CT images were input into 2D U-plus-net.The obtained model was trained and tested by 5-fold cross-validation method.Finally,the accuracy of automatic heart segmentation was evaluated by Dice coefficient,HD95 and mean surface distance.Results The Dice coefficient,mean surface distance and HD95 of automatic heart segmentation on Data A were 0.941±0.012,(3.918±0.201)mm and(5.863±0.561)mm,respectively,while those of automatic heart segmentation on Data B were 0.934±0.014,(4.112±0.320)mm and(6.035±0.659)mm,respectively.Conclusion The automatic segmentation method based on 2D/3D U-plus-net improves the accuracy of automatic heart segmentation.