Projection domain and image domain joint learning reconstruction network for reconstructing chest limited angle CT images
Objective To observe the value of dual domain(projection domain and image domain)joint learning reconstruction network(DDRNet)for reconstructing chest limited angle CT images.Methods Totally 4 300 chest enhanced CT images of 65 patients with chest tumors were retrospectively enrolled and reconstructed with DDRNet,and 3D and 2D projection information fusion were performed.The reconstruction effect of DDRNet was evaluated and compared with that of single domain reconstruction and filtered back projection(FBP),residual encoder-decoder convolutional neural network(RED-CNN),Resnet and deconvolution network(RDN),as well as of generative adversarial network(GAN).Results The peak signal to noise ratio(PSNR)of DDRNet reconstructed images tended to stabilize after approximately 60 iterations,while the projection domain and image domain learning networks tended to stabilize after approximately 90 and 80 iterations.After stable training,compared to the projection domain learning network,the fluctuation of output results of DDRNet and image domain learning networks were less.After 200 rounds of training,PSNR of DDRNet reconstructed images was significantly higher than that of projection domain and image domain learning networks.The quality of DDRNet reconstructed image was significantly better than that of FBP,RED-CNN,RDN and GAN.Conclusion DDRNet could be used to effectively reconstruct high-quality chest limited angle CT images.
万方数据
维普
