摘要
针对深度学习训练模型过程中存在感受野小及特征丢失等问题,提出一种基于生成对抗网络的心脏核磁共振图像分割网络MCA GAN,提取心肌内外膜的同时保留更多的浅层信息和深层信息,提高分割精度。将MCA GAN在多个心脏MRI分割竞赛的公开数据集上进行训练,并与FCN和DCGAN两个神经网络进行实验对比。实验结果表明,相较于FCN和DCGAN,MAC GAN的Dice系数内外膜提升了1.44%和3.18%,Jaccard系数提升了2.12%和3.35%,Sensitivity系数提升了1.18%和1.80%,和其他方法相比较均有显著提升。
To address the problems of small perceptual field and feature loss during deep learning training models,a genera⁃tive adversarial network⁃based cardiac MRI image segmentation network,MCA GAN,is proposed to extract the inner and outer myocardial membranes while retaining more superficial and deep information to improve segmentation accuracy.The MCA GAN is trained on public datasets of several cardiac MRI segmentation competitions and experimentally compared with two neural net⁃works FCN and DCGAN.The experimental results show that compared with FCN and DCGAN,MAC GAN has improved Dice coef⁃ficient by 2.5%and 2.7%,Jaccard coefficient by 2.0%and 2.7%,and Sensitivity coefficient by 2.5%and 1.8%,all of which are sig⁃nificantly improved comparing with other methods.
作者
李孟歆
韩煜
李松昂
贾欣润
李易营
Li Mengxin;Han Yu;Li Song’ang;Jia Xinrun;Li Yiying(School of Electrical and Control Engineering,Shenyang Construction University,Shenyang 110168)
出处
《现代计算机》
2023年第3期1-8,共8页
Modern Computer
基金
辽宁省自然科学基金(20180550059)。
关键词
多尺度结构
综合注意力
生成对抗网络
左心室分割
multi⁃scale
attentional mechanism
generative adversarial networks
left ventricular segmentation