基于多级串联深度卷积神经网络配准大形变图像

Registration of large deformation images based on multiple cascaded deep convolutional neural networks

目的采用无监督方式基于多级串联深度卷积神经网络(CNN)建立大形变图像配准网络(LDIRnet)模型,评估其配准脑部MRI及肺部CT图像的性能。方法串联多个结构相同而参数不同的深度CNN,以端到端方式学习待配准图像之间的多个小形变场;再通过叠加小形变场计算待配准图像之间的大形变场,实现大形变图像配准。结果配准3D脑部MRI时,三级LDIRnet(Three-LDIRnet)配准性能最佳,Dice系数为0.793±0.104,其次为两级LDIRnet(Two-LDIRnet)及ANT,VoxelMorph最差;配准肺部4D CT图像时,总体配准性能以mlVIRNET最佳,平均配准误差为(1.84±1.39)mm,其次为Three-LDIRnet、DLIR及Two-LDIRnet,VoxelMorph模型误差最大。结论多级串联配准网络模型有助于提高配准大形变图像的精度。

Objective To establish a large deformation images registration network(LDIRnet)model with unsupervised method based on multiple cascaded deep convolution neural networks(CNN),and to evaluate its performance for the registration of brain MRI and lung CT images.Methods Serial small deformation fields were learned end-to-end via LDIRnet,which was composed of serial convolutional networks with same structures and different learnt weights.A final large deformation field was obtained through adding all small deformation fields,so as to realize registration of large deformation images.Results For the registration of 3D brain MR images,Three-LDIRnet had the best registration performance,with Dice coefficient of 0.793±0.104,followed by Two-LDIRnet and ANT,and VoxelMorph was the worst.For the registration of 4D lung CT images,mlVIRNET had the best registration performance,with an average registration error of(1.84±1.39)mm,followed by Three-LDIRnet,DLIR and Two-LDIRnet.The largest error occurred in VoxelMorph model.Conclusion Multiple cascaded registration network model was helpful to improve the accuracy of registration of large deformation images.